fix spurious recompilation on incremental editable builds

[matthiasdiener]

Description

Two unrelated causes made pip install -e . recompile from scratch even
with no source changes:

• PyTorch extension: setuptools' editable_wheel points build_ext.build_temp
  at an ephemeral /tmp/*.build-temp dir, so all csrc objects recompiled every
  build. Override finalize_options() to pin build_temp to build/torch_temp,
  giving ninja a persistent object cache.

• CK grouped gemm: ck_grouped_gemm_common.h had a malformed relative include
  ("../../common/util/system.h", which resolves to a nonexistent path). This
  forced hipify's output-dir-dependent fallback resolution, so the pytorch and
  common hipify passes emitted different include paths and the generated
  _hip.h flip-flopped every build, retriggering all 8 CK objects. Correct the
  include to "../../util/system.h" so both passes agree.

Companion to #666 and Micky774/QoLA#9.
With all three PRs, total rebuild time without source code changes is reduced from ~8 minutes to ~6 seconds, with zero recompilations.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[github-actions]

Nit: naming — torch_temp reads as PyTorch-specific, but get_build_ext() is also called from transformer_engine/jax/setup.py:60, so JAX ext objects will land under build/torch_temp/ too. Since both frameworks share this dir, a framework-neutral name like ext_temp (or framework_temp) would be less confusing when someone stumbles into it later. No functional impact — ninja tracks objects by source path so pytorch/jax objects won't collide.

[github-actions]

Reviewed the two-file diff. Both fixes look correct and targeted:

• build_tools/build_ext.py: finalize_options override pinning build_temp to build/torch_temp correctly addresses the ephemeral setuptools editable_wheel path. build/ is already gitignored. Left one non-blocking nit on the directory name (used by JAX ext too).
• ck_grouped_gemm_common.h: Include path fix is verified — transformer_engine/common/util/system.h exists; the previous ../../common/util/system.h resolved to a nonexistent transformer_engine/common/common/util/system.h. Note: the file now includes the same header via two paths (line 19 common/util/system.h and line 21 ../../util/system.h) but that's harmless with #pragma once and matches the existing dual-style pattern in the file.

Copyright headers: OK (both files have current AMD end-year 2026; NVIDIA line preserved).

Verdict: LGTM, already approved by @Micky774.