[PyTorch debug] Fixes to debug tests failures

[pggPL]

Description

This PR fixes 2 issues with debug tests:

• distributed tests for debug were failing on Ampere because test_log in test_distributed needs fp8, but it was not skipped on non fp8 devices,
• The test_perf used to check if the CPU time of debug layers (with no features active) matched non-debug layers. This test was unstable and sometimes failed. I replaced it with a deterministic test that verifies inspect_tensor_enabled is called only once when the feature is never used. Previously, calling this API every iteration caused heavy CPU overhead. That issue was fixed in [PyTorch debug] Improve precision debug tools performance #1909.

I added option inspect_only_once to TestDummyFeature - in this case inspect_tensor_enabled will return True, None - which means "invoke inspect_tensor() this iteration, but the future will be never invoked again".

Type of change

• test fix

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

[pggPL]

/te-ci pytorch L1

[pggPL]

/te-ci pytorch