Disable Flash Attention with USE_FLASH_ATTENTION

[alvarobartt]

What does this PR do?

This PR handles the value for the environment variable USE_FLASH_ATTENTION when applicable, so that Flash Attention can be disabled, as it's known to be problematic on Turing.

This being said, even if the USE_FLASH_ATTENTION environment variable was already there, it was not handled for some cases such as Qwen2, Qwen3, or GTE; neither for Mistral which won't support it, but still raising an error if provided.

Fixes #644 #347

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the documentation guidelines, and here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@Narsil

Additional notes

Maybe it's a nice time to update the documentation around this issue, as there are similar reports of numerical instability / null or empty vectors when running on Turing, which is expected, but maybe the documentation could be more clear around it if so many users report the same issues; which I can do as part of this PR if needed 🤗