Introduce a fused padding + cast transpose kernel grouped linear

[alextmagro]

Fuses the following 2 kernels:

Pad tensors in BF16
Cast from BF16 to FP8, transpose, and store

into a single kernel that:

Cast from BF16 to FP8, transpose, store with padding

2x speedup over unfused kernel, applicable to grouped linear.

[github-actions]

Claude code review summary

Reviewed PR #632 (fused padding + cast-transpose for grouped linear) against dev (HEAD^1). Scope: 11 files, ~640 LoC added across the C API, ROCm kernel, PyTorch extension, Python GroupedLinear, C++ test, Python test, and benchmark.

Overall: Approach is sound — a single padded MCT kernel replacing pad-then-MCT, exposed end-to-end through split_quantize(valid_split_sections=...) and consumed by GroupedLinear(actual_m_splits=, unpad_output=). The ROCm kernel padding semantics (input zero-fill, amax/store gated by valid_num_rows, columnwise store gated by num_rows) are internally consistent. Upstream isolation is good — almost all new logic sits behind #ifdef __HIP_PLATFORM_AMD__ or IS_HIP_EXTENSION.

Findings (see inline comments):

• Blocker — likely compile failure in the new performTestWithPadding C++ test: references valid_h that was never declared (the loop binds in_h).
• One unguarded change to shared CUDA-reachable code in multi_cast_transpose.cu (NVTE_CHECK operand swap) — semantically equivalent on CUDA but creates IFU noise. Suggest guarding or commenting.
• GroupedLinear.forward signature gains actual_m_splits / unpad_output without docstring updates; the dim-0 view at return assumes the unpadded case.
• A few small Python nits (stray \ in a comment, mutation of grad_output_quantizers without explanation).

Copyright headers: OK — all 11 modified files have correctly updated AMD headers with the 2026 end-year and preserved NVIDIA lines.

Not duplicated: @ipanfilo's nit on test_multi_cast_transpose.cu:108 about marking the make_nvte_vector lambda removal with ROCm: for IFU clarity still stands.

[alextmagro]

Addressed copilot comments

[ipanfilo]

Pending CI results