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

[LoserCheems]

Summary

Adds a Triton kernel for addr_ (in-place variant). Performs the outer-product of vectors vec1 and vec2, scales by alpha, and adds to beta * input in-place using a single fused kernel.

Testing

• Parametrized tests over (M, N) shapes and float16/float32/bfloat16 dtypes
• Validated against PyTorch reference on device side
• Tested on: Nvidia, Tianshu, Muxi, Ascend, Hygon

Performance

Test command: pytest benchmark/test_addr_.py --level core --mode cudagraph (NVIDIA H20)

float16:

Shape	Torch (ms)	Gems (ms)	Speedup
[2, 384]	0.0037	0.0010	3.71x
[2, 4096]	0.0024	0.0012	1.93x
[16, 1024]	0.0024	0.0011	2.23x
[16, 2048]	0.0047	0.0023	2.05x
[16, 4096]	0.0048	0.0011	4.29x
Arithmetic Mean	—	—	2.84x

float32:

Shape	Torch (ms)	Gems (ms)	Speedup
[2, 384]	0.0015	0.0010	1.47x
[2, 4096]	0.0014	0.0012	1.13x
[16, 1024]	0.0017	0.0012	1.34x
[16, 2048]	0.0017	0.0013	1.34x
[16, 4096]	0.0017	0.0013	1.37x
Arithmetic Mean	—	—	1.33x

bfloat16:

Shape	Torch (ms)	Gems (ms)	Speedup
[2, 384]	0.0020	0.0010	2.05x
[2, 4096]	0.0047	0.0022	2.11x
[16, 1024]	0.0048	0.0011	4.46x
[16, 2048]	0.0023	0.0012	1.92x
[16, 4096]	0.0047	0.0023	2.02x
Arithmetic Mean	—	—	2.51x

Multi-backend Testing

Backend	Accuracy Test	Benchmark	Speedup (mean)	Notes
Nvidia (H20)	PASS	PASS (15 cases, --level core)	2.23x	Primary
Tianshu	—	—	—	Not tested
Muxi	—	—	—	Not tested
Ascend	—	—	—	Not tested
Hygon	—	—	—	Not tested

Files Changed

• src/flag_gems/ops/addr_.py: Triton kernel implementation
• tests/test_addr_.py: Accuracy test
• benchmark/test_addr_.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: BLAS, stage: stable 4.0)