[Common] Skip cuDNN 9.10.0/9.10.1 due to bugs

[cyanguwa]

Description

cuDNN fixed a few bugs in cuDNN 9.10.2 and this PR is to:

• skip cuDNN 9.10.0 for SDPA FP8 related bugs,
• skip cuDNN 9.10.0 and 9.10.1 for SDPA FP16/BF16 related bugs,
• unify the use of ModelConfig in all unit tests, and
• fix FP8 and certain FP16/BF16 tests so they check backend availability first before running any configs.

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 see description.

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

[cyanguwa]

/te-ci pytorch L1 L0

[cyanguwa]

Pipeline 31306522 for 9.10.0 + 25.06 + 12.9
Pipeline 31306722 for 9.10.1 + 25.06 + 12.9

[timmoon10]

test_fused_attn.py is kind of a strange place to keep ModelConfig. It's indicating that these tests are becoming over-tuned to attention. We could avoid this refactor by moving test_sanity_attention_extra_state into the fused attention tests:

TransformerEngine/tests/pytorch/test_sanity.py

Line 1119 in 7b249ae

def test_sanity_attention_extra_state(model, dtype):

Alternatively, we could move ModelConfig from test_fused_attn.py into utils.py. It has many attention-specific options though, so I'd prefer having separate implementations.

[zhongbozhu]

LGTM