Fix float out of range in owlvit and owlv2 when using FP16 or lower precision

[aliencaocao]

What does this PR do?

As discussed in #31342 (comment)

The hard coded -1e6 value used to mask logits is not in range for FP16 which causes a value cannot be converted to type at::Half without overflow error. We replace that with the min value of whichever dtype the logits are in.

Did a search and only the owlvit and owlv2 models are using this, however, we should take note if any future model implementation hardcodes a similar value. There exists a couple more models using -1e7, -1e8, -1e9 and -1e10, but most of them are TensorFlow models, and the rest are either deprecated or language models. Since there has been no issue reports on this so far, I think it is properly handled somewhere so I won't change them.

Tests on FP32 showed no numerical difference in logits up to 1e-4 (what the test uses).

The reason of not using float('-inf') is that it can be 20% (RTX 3080Ti) to 3x (H100 SXM5) slower than using a predefined value for some reason.

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?

@amyeroberts

[amyeroberts]

Thanks for fixing and for investigating across the models in the library ❤️

[aliencaocao]

@amyeroberts Anything else to do before it can be merged? It would fix the last failing test for #31342

[amyeroberts]

Nope! I'll merge now. Thanks for pinging and reminding me, and sorry for the delay!