[PyTorch][CI] Connect test files that are not wired into qa, fix the broken one

[pggPL]

Description

Six pytorch test files are not referenced by any script in qa/, so they have never run in CI.
git log -S<name> --all -- qa/ shows that none of them was ever added and later removed, and none
of the PRs that introduced them touched qa/ — they were simply never wired up.

file	status on current main
test_qk_norm.py	45 passed
test_float8_current_scaling_exact.py	5 passed
attention/test_cu_seqlens_cache.py	1 passed, 1 skipped (needs 2 GPUs)
test_nvfp4_fsdp2_hooks.py	16 skipped (needs sm_100+)
test_fused_router_perf.py	5 skipped (gated behind TE_RUN_PERF_TESTS)
layernorm_mlp/test_selective_activation_checkpoint.py	16 failed

This PR connects four of them and fixes the one that turned out to be broken. All are single-GPU
and self-skip on unsupported hardware. Added L0 time is roughly 13 s.

test_fused_router_perf.py is left out — it is gated behind TE_RUN_PERF_TESTS and deserves its
own decision.

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

• Add L0 entries for test_qk_norm.py, test_float8_current_scaling_exact.py,
  attention/test_cu_seqlens_cache.py and layernorm_mlp/test_selective_activation_checkpoint.py.
• Move test_nvfp4_fsdp2_hooks.py into tests/pytorch/nvfp4/, which L0 already runs as a whole
  directory.
• Also run attention/test_cu_seqlens_cache.py in L1, where its cross-device test can actually run
  instead of being permanently skipped.
• Fix test_selective_activation_checkpoint.py. The feature is fine — outputs and all six
  parameter gradients are bit-exact between the checkpointed and non-checkpointed path. The test
  asserted a 6x forward memory ratio that the model shape in the test can never reach; it now
  asserts on the memory that recompute actually frees. Also drops a flaky wall-clock assertion,
  checks correctness before memory, and skips configurations that do not fit in device memory.

The commit messages carry the full derivation and measurements.

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

[greptile-apps]

Greptile Summary

This PR connects previously uncollected PyTorch tests to CI and repairs the selective activation-checkpointing test.

• Adds four single-GPU test files to the L0 PyTorch job.
• Adds the cross-device sequence-length cache test to the L1 distributed job.
• Replaces an unreachable checkpointing memory-ratio assertion with a configuration-derived savings check and skips configurations that exceed available memory.
• Moves the NVFP4 FSDP2 hook test into the NVFP4 directory already collected by L0.

Confidence Score: 5/5

The PR appears safe to merge, with the newly wired tests correctly collected and no actionable regressions identified.

The QA invocations match existing orchestration patterns, the multi-device cache test works through ordinary visible CUDA devices, the moved NVFP4 test has no location-sensitive dependencies, and the revised checkpointing assertions reflect the configured activation allocations.

Important Files Changed

Filename	Overview
qa/L0_pytorch_unittest/test.sh	Adds direct pytest invocations for four previously uncollected single-GPU test files with unique JUnit outputs and existing failure handling.
qa/L1_pytorch_distributed_unittest/test.sh	Runs the sequence-length cache test with multiple visible GPUs so its cross-device case can execute without distributed initialization.
tests/pytorch/layernorm_mlp/test_selective_activation_checkpoint.py	Replaces a structurally unreachable memory-ratio threshold, prioritizes numerical checks, and adds a conservative free-memory skip guard.
tests/pytorch/nvfp4/test_nvfp4_fsdp2_hooks.py	Moves an unchanged, path-independent test into the NVFP4 directory already collected by the L0 job.

_{Reviews (1): Last reviewed commit: "[CI] Connect orphaned pytorch test files..." | Re-trigger Greptile}