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

[chx7514]

Summary

Adds a Triton kernel for _fake_quantize_learnable_per_channel_affine_backward. Triton kernel implementation for _fake_quantize_learnable_per_channel_affine_backward,
the backward of learnable per-channel fake quantization. Computes the gradients
with respect to the input, scale and zero-point, accounting for the saturation
branch where the quantized value is clamped to [quant_min, quant_max].

Testing

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

Performance

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

fake_quantize_learnable_per_channel_affine_backward

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float32	[64, 1024]	0.074368	0.062336	1.193
float32	[1024, 64]	0.078336	0.072224	1.085
float32	[256, 512, 32]	0.144576	0.251744	0.574
float32	[128, 256, 128]	0.182240	0.246272	0.740
float32	[64, 1024]	0.071600	0.063520	1.127
float32	[4096, 4096]	0.309472	0.219392	1.411
float32	[16, 128, 64, 60]	0.254128	0.279744	0.908
float32	[1024, 1024]	0.076160	0.065024	1.171

Operator	Geometric Mean Speedup
fake_quantize_learnable_per_channel_affine_backward	0.991

Multi-backend Testing

Backend	Accuracy Test	Speedup (mean)	Notes
Nvidia (H20)	PASS (8 cases)	0.991	Primary
Tianshu	N/A	—	—
Muxi	N/A	—	—
Ascend	N/A	—	—
Hygon	N/A	—	—

Files Changed

• src/flag_gems/ops/_fake_quantize_learnable_per_channel_affine_backward.py: Triton kernel implementation
• tests/test_fake_quantize_learnable_per_channel_affine_backward.py: Accuracy test
• benchmark/test_fake_quantize_learnable_per_channel_affine_backward.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: Math, stage: alpha 5.4)

[bin913]

/test|_fake_quantize_learnable_per_channel_affine_backward:h20