[Triton] Add Fused GEMM A8W8 + Split + Concat Triton Kernel

[farlukas]

Motivation

This PR adds a new Triton fusion kernel. The fused kernel performs an FP8 GEMM using a blockscale quantization approach. It will then split the GEMM result into two 3D tensors. Finally, it will concatenate a third tensor to the first split of the GEMM result.

Technical Details

Equivalent to the following sequence:

c = (x @ w).view(-1, y.shape(1), S1 + S2)
c1, c2 = c.split([S1, S2], dim=-1)
c1 = c1.cat(y, dim=-1)
return c1, c2

Test Plan

pytest pytest op_tests/triton_tests/test_fused_gemm_a8w8_blockscale_split_cat.py

Test Result

Passes all unit tests comparing Triton output with PyTorch output.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.