Fix Gemma2 dtype issue when storing weights in float16 precision

[Nerogar]

What does this PR do?

This change fixes an exception that can happen when storing Gemma2 weights in float16 format, but using a bfloat16 autocast context. In that case, hidden_states can be in bfloat16 format, while attention_mask is in float16 format. That will lead to the following error

    attention_mask = torch.where(sliding_window_mask, min_dtype, attention_mask)
RuntimeError: value cannot be converted to type at::Half without overflow

Using the min value of the attention_mask instead should be safe, since the value is only used in the torch.where operation.

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?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[LysandreJik]

Seems reasonable to me, cc @ArthurZucker