remove deprecated tests

[jiqing-feng]

This PR removes deprecated test coverage to reduce noise and avoid generating cases that are intentionally unsupported.

Changes

• tests/test_autograd.py
  • Stop generating req_grad[1] == True for test_matmullt.
  • Remove unreachable/dead assertions tied to that deprecated path.
• tests/test_linear8bitlt.py
  • Restrict test_linear_serialization to non-deprecated has_fp16_weights=False.
  • Remove skip-based handling for deprecated has_fp16_weights=True combinations.
• tests/test_modules.py
  • Remove deprecated test cases.
• tests/test_functional.py
  • Remove deprecated IGEMM functional test block.

Impact

• Smaller and cleaner test matrix.
• No behavior change in supported runtime paths.
• CI signal improves by removing deprecated/skip-heavy combinations.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

This is the main one where I hesitate because we may still have external users calling int8_double_quant - even if they're not actually training int8 weights.

There's a tip in the docs about this but this is one where we haven't emitted deprecation warnings on either.

<Tip>
    This function is useful for training, but for inference it is advised to use [`int8_vectorwise_quant`] instead.
    This implementation performs additional column-wise transposed calculations which are not optimized.
</Tip>

My ask here is that for now we change this one to only test on CUDA and give this one a short deprecation cycle before removing. This test doesn't necessarily assume int8 training and I believe is our only coverage on the op.

[matthewdouglas]

I'm not fundamentally opposed to removing these tests as we generally either don't run them, or would be unlikely to take action on them if they did fail. The main hesitation is that we haven't gone through a deprecation cycle on some of these. Particularly int8_double_quant as mentioned.

As for the rest, considering I was planning to start emitting deprecation warnings for all of those features specifically before we push out v0.50.0, I think they are OK.

[jiqing-feng]

Hi @matthewdouglas . I have added test_int8_double_quant back, and only limit it on cuda. Please review it. Thanks!

.