[WIP] N*G Triton group gemm for MoE

[lessw2020]

This PR adds a Triton Group GEMM with full backwards pass support, for integration with MoE training.
The forward pass is from FBGemm experimental:
https://github.com/pytorch/FBGEMM/blob/main/fbgemm_gpu/experimental/gemm/triton_gemm/grouped_gemm.py

1 - Numerics with BF16 have been verified on sample sizes and the core DeepSeek v3 shapes.

2025-03-13 16:40:08,294 - INFO - Gradient shapes - grad_x: torch.Size([1024, 256]), grad_w: torch.Size([10240, 256])
2025-03-13 16:40:08,294 - INFO - Running PyTorch reference implementation
2025-03-13 16:40:08,591 - INFO - Comparing gradients with PyTorch reference
2025-03-13 16:40:08,602 - INFO - Maximum gradient error - grad_x: 0.125, grad_w: 0.125
2025-03-13 16:40:08,641 - INFO - ✓ SUCCESS! grad_X matches the PyTorch reference (allclose check passed)
2025-03-13 16:40:08,641 - INFO - ✓ SUCCESS! grad_W matches the PyTorch reference (allclose check passed)
2025-03-13 16:40:08,641 - INFO - Gradients allclose check - grad_x: True, grad_w: True
2025-03-13 16:40:08,641 - INFO - ✓ SUCCESS: Gradients match the PyTorch reference (allclose check passed)

2 - Todos:
a - fp8 support
b - TMA (was removed to focus on numerics)
c - WS
d - Perf and auto-tuning

3 - Integration - ready now for BF16 though may want to do perf work first.