[CPU/CUDA EP] Add DeformConv op support (#27393)

### Description
<!-- Describe your changes. -->
This change adds support for the Deformable Convolution 2D operator
(DeformConv2D) to ONNX Runtime. The branch implements the operator
schema and registration, provides kernel implementations (CPU and
GPU/CUDA where available), implements shape inference, and adds unit and
integration tests to validate correctness and numerical parity with
reference implementations. The changes include performance-oriented
optimizations and necessary changes to build/test scripts.


### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

Deformable convolutions are widely used in vision models that require
spatial sampling flexibility (e.g., Deformable ConvNets, some
detection/segmentation models). Native support in ONNX Runtime enables
these models to run efficiently without custom operators or external
runtimes, broadening the set of compatible models and improving
performance and portability.


### See also

- https://onnx.ai/onnx/operators/onnx__DeformConv.html
-
https://docs.pytorch.org/vision/main/generated/torchvision.ops.deform_conv2d.html
- https://arxiv.org/abs/1811.11168
- https://arxiv.org/abs/1703.06211
-
https://github.com/pytorch/vision/blob/0f6d91d9fe514e6de2f5519114cbeb389d498b2d/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu
-
https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp
-
https://github.com/pytorch/vision/blob/0f6d91d9fe514e6de2f5519114cbeb389d498b2d/torchvision/csrc/ops/cpu/deform_conv2d_kernel.cpp
-
https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp

- #22060
- #15572
- #20810
- #16903
- onnx/onnx#5451
- ZhengPeng7/BiRefNet#167
- pytorch/pytorch#68910
- pytorch/vision#2066

Author: ShirasawaSama

docs/OperatorKernels.md

@@ -103,6 +103,8 @@ Do not modify directly.*
 |||[11, 13]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64)<br/> **T2** = tensor(int32), tensor(int64)|
 |DFT|*in* input:**T1**<br> *in* dft_length:**T2**<br> *in* axis:**tensor(int64)**<br> *out* output:**T1**<br><br>or<br><br>*in* input:**T1**<br> *in* dft_length:**T2**<br> *out* output:**T1**|20+|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int32), tensor(int64)|
 |||[17, 19]|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int32), tensor(int64)|
+|DeformConv|*in* X:**T**<br> *in* W:**T**<br> *in* offset:**T**<br> *in* B:**T**<br> *in* mask:**T**<br> *out* Y:**T**|22+|**T** = tensor(double), tensor(float)|
+|||[19, 21]|**T** = tensor(double), tensor(float)|
 |DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(uint8)|
 |||[11, 12]|**T** = tensor(double), tensor(float), tensor(uint8)|
 |||[1, 10]|**T** = tensor(double), tensor(float)|
@@ -697,6 +699,8 @@ Do not modify directly.*
 |Crop|*in* input:**T**<br> *out* output:**T**|1+|**T** = tensor(double), tensor(float), tensor(float16)|
 |CumSum|*in* x:**T**<br> *in* axis:**T2**<br> *out* y:**T**|14+|**T** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T2** = tensor(int32), tensor(int64)|
 |||[11, 13]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T2** = tensor(int32), tensor(int64)|
+|DeformConv|*in* X:**T**<br> *in* W:**T**<br> *in* offset:**T**<br> *in* B:**T**<br> *in* mask:**T**<br> *out* Y:**T**|22+|**T** = tensor(bfloat16), tensor(double), tensor(float), tensor(float16)|
+|||[19, 21]|**T** = tensor(double), tensor(float), tensor(float16)|
 |DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[11, 12]|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[1, 10]|**T** = tensor(double), tensor(float), tensor(float16)|

onnxruntime/core/providers/cpu/cpu_execution_provider.cc

@@ -1220,6 +1220,8 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain,
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, uint8_t, Resize);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 20, Scan);
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 20, Shape);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 21, float, DeformConv);
+class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 21, double, DeformConv);
 
 // Opset 20
 class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 20, 20, ConstantOfShape);
@@ -1316,6 +1318,8 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Ac
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Atanh);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Conv);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, ConvTranspose);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, float, DeformConv);
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, double, DeformConv);
 class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Det);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, float_float, Dropout);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, float_double, Dropout);
@@ -3277,6 +3281,8 @@ Status RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
                                                                   Resize)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 20, Scan)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 20, Shape)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 21, float, DeformConv)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 19, 21, double, DeformConv)>,
 
       // Opset 20
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 20, 20,
@@ -3407,6 +3413,8 @@ Status RegisterOnnxOperatorKernels(KernelRegistry& kernel_registry) {
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Atanh)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Conv)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, ConvTranspose)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, float, DeformConv)>,
+      BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, double, DeformConv)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, Det)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, EyeLike)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 22, GlobalAveragePool)>,