[detorch] remove vestigial py_itfs_common.h includes (non-CK); top_k_per_row now torch-free

[amd-ruitang3]

Summary

Removes every vestigial #include "py_itfs_common.h" from the non-CK TUs
that include it but reference none of its exported symbols. That header carries
an unguarded #include <torch/all.h>, so a stray include silently drags the whole
torch/ATen header tree into a module's JIT build.

Two kinds of benefit:

1. Immediate torch-free win — module_top_k_per_row. Its kernel TU was
   de-torched in [module_topk_*] de-torch topk_per_row / topk_plain + externalize workspaces #4702, but the two ASM sibling TUs
   (asm_topk_per_row_{decode,prefill}.cu) were missed — each included only
   aiter_tensor.h + py_itfs_common.h, so the header was their sole torch
   source. Removing it makes the entire module closure (13 headers) torch-free.
2. Latent-regression removal — the rest. asm_mha_{fwd,bwd,varlen_fwd,varlen_bwd}.cu
   and fused_ar_mhc_post.cu keep their own direct torch includes (they are still
   torch-bound), so this is not a compile-time win today. But it deletes the exact
   trap that left topk half-de-torched: once those modules are de-torched later, a
   leftover py_itfs_common.h would silently re-introduce <torch/all.h>.

Why each removal is safe

file	py_itfs_common symbols used	how it still gets what it needs
asm_topk_per_row_{decode,prefill}.cu	none	aiter_tensor.h (already included, line 3) pulls aiter_hip_common.h → AITER_CHECK/HIP_CALL/HipDeviceGuard + HIP runtime; now torch-free
asm_mha_{fwd,bwd,varlen_fwd,varlen_bwd}.cu	none	torch/all.h + ATen/hip/HIPContext.h remain (load-bearing); getCurrentHIPStream here is at::hip::, so no aiter_hip_common symbol was in play
fused_ar_mhc_post.cu	none	already #include "aiter_hip_common.h" directly (line 8); ATen/c10 includes remain

py_itfs_common.h's torch-typed exports (torch_fp8 / torch_fp4x2 / t2ck /
torchDTypeToStr) are used only by CK TUs that need torch at the boundary anyway;
those includes are left untouched (out of scope).

Validation (in-container, gfx942 / ROCm 7.2.3)

• module_top_k_per_row — fresh JIT rebuild from a cleared build dir compiles clean; recursive include closure = 0 torch/ATen/c10 headers.
  • op_tests/test_topk_per_row.py (decode): all shapes match torch.topk; [mb_workspace_reuse] PASS.
  • op_tests/test_topk_row_prefill.py (prefill): standard and fast kernels Passed, all_close_standard/all_close_fast == True.
• module_fmha_v3_fwd — fresh rebuild clean (39.4s). asm_mha_fwd.cu is include-identical to the other three asm_mha TUs.
• module_fused_ar_mhc — fresh rebuild clean (36.5s).

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4756 --add-label <label>