Fix fp8 UT bugs

[YangKai0616]

What does this PR do?

When running the tests with the command pytest -rA tests/quantization/finegrained_fp8/test_fp8.py, the current output reports the following errors:

FAILED tests/quantization/finegrained_fp8/test_fp8.py::FP8QuantizerTest::test_quantized_model - AssertionError: 'Once upon a time, there was a man who was a farmer.' != 'Once upon a time, there was a man who was very rich.'
FAILED tests/quantization/finegrained_fp8/test_fp8.py::FP8QuantizerTest::test_quantized_model_conversion - AssertionError: 122 != 146
FAILED tests/quantization/finegrained_fp8/test_fp8.py::FP8QuantizerTest::test_quantized_model_multi_accelerator - AssertionError: 'Once upon a time, there was a man who was a farmer.' != 'Once upon a time, there was a man who was very rich.'
FAILED tests/quantization/finegrained_fp8/test_fp8.py::FP8QuantizerTest::test_save_pretrained - AssertionError: 'Once upon a time, there was a man who was a farmer.' != 'Once upon a time, there was a man who was very rich.'
FAILED tests/quantization/finegrained_fp8/test_fp8.py::FP8QuantizerTest::test_save_pretrained_offload - AssertionError: 'Once upon a time, there was a man who was a farmer.' != 'Once upon a time, there was a man who was very rich.'

The errors can be categorized into two types:

1. Model output does not match EXPECTED_OUTPUT

Root cause: The _dtype used meta model parameter dtype in here, while the EXPECTED_OUTPUT appears to be the result obtained with a fixed dtype of torch.float32. After I explicitly set _dtype to torch.float32, the result matches the expected output.

This PR updates the EXPECTED_OUTPUT accordingly. Please let me know your insights, thx!

2. nb_linears - 25 ≠ nb_fp8_linear

Root cause: The replacement operation in here can lead to patterns like XX.fc1.XX → XX.fc*.XX, in which case the layer name fc1 listed in modules_to_not_convert fails to match the actual layer.

This PR fixes the string matching logic.

[YangKai0616]

@SunMarc , @ydshieh , please help review, thanks!

[SunMarc]

Thanks ! Left a comment

[SunMarc]

maybe we can use the same logic as mxfp4 ? We have the following function should_convert_module. Indeed here the logic here changed recently and this needs to be fixed

[YangKai0616]

Done.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: finegrained_fp8

[SunMarc]

This is a bit more complex to fix those due to the refactor, so I decided to open a PR to fix those. I've added you as a co-author. Thanks for running those tests ;)

[YangKai0616]

Alright, I will follow the changes in the new PR and close this PR.