[KernelGen][Nvidia] Add transpose_ operator with Triton kernel

[ShawnsYing]

Summary

Adds a Triton kernel for transpose_. In-place transpose: swaps two dimensions of the tensor, mutating it in place
and returning it. This is a zero-copy view operation that swaps the sizes and
strides of the two specified dimensions via as_strided, matching the semantics
of aten::transpose_.

Testing

• Validated against reference on device via to_reference(inp, True)
• Tested on: Nvidia, Tianshu, Muxi, Ascend, Hygon

Performance

Test command: pytest benchmark/test_transpose_.py --level core (NVIDIA H20)

transpose_ (in-place)

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float16	[64, 64]	0.003424	0.003200	1.070
float16	[256, 512]	0.003424	0.003264	1.049
float16	[1024, 1024]	0.003328	0.003232	1.030
float16	[4096, 4096]	0.003232	0.003232	1.000
float16	[64, 512, 512]	0.003232	0.003296	0.981
float16	[128, 256, 64]	0.003328	0.003296	1.010
float16	[8, 16, 32, 64]	0.003232	0.003232	1.000
float32	[64, 64]	0.003232	0.003296	0.981
float32	[256, 512]	0.003264	0.003264	1.000
float32	[1024, 1024]	0.003264	0.003248	1.005
float32	[4096, 4096]	0.003264	0.003264	1.000
float32	[64, 512, 512]	0.003264	0.003232	1.010
float32	[128, 256, 64]	0.003232	0.003264	0.990
float32	[8, 16, 32, 64]	0.003264	0.003264	1.000
bfloat16	[64, 64]	0.003264	0.003264	1.000
bfloat16	[256, 512]	0.003264	0.003264	1.000
bfloat16	[1024, 1024]	0.003264	0.003232	1.010
bfloat16	[4096, 4096]	0.003264	0.003264	1.000
bfloat16	[64, 512, 512]	0.003264	0.003296	0.990
bfloat16	[128, 256, 64]	0.003264	0.003264	1.000
bfloat16	[8, 16, 32, 64]	0.003264	0.003264	1.000

Operator	Arithmetic Mean Speedup
transpose_ (in-place)	1.006

Multi-backend Testing

Backend	Accuracy Test	Speedup (mean)	Notes
Nvidia (H20)	PASS (21 cases)	1.006	Primary
Tianshu	N/A	—	—
Muxi	N/A	—	—
Ascend	N/A	—	—
Hygon	N/A	—	—

Files Changed

• src/flag_gems/ops/transpose_.py: Triton kernel implementation
• tests/test_transpose_.py: Accuracy test
• benchmark/test_transpose_.py: Performance benchmark
• src/flag_gems/ops/__init__.py: Register import and __all__
• src/flag_gems/__init__.py: Register to _FULL_CONFIG
• conf/operators.yaml: Add operator entry (kind: Tensor, stage: alpha 5.4)