Fix Windows and onnx dtype compatibility

[IlyasMoutawwakil]

What does this PR do?

Fixes multiple windows specific issues:

• dtype mismatch: numpy uses a default int representation that depends on os. On ubuntu/macos it uses int64 but on windows it uses int32. This results in tokenizers returning input ids and attention masks in a different format than torch's default which is int64/long.
• shutil.rmtree doesn't always work on windows, there's some random failures that we can see when a tempfile is being closed (cleaned up) or when directly trying to shutil.rmtree a folder.

Also seq2seq decoder used to return past kv cache in torch format all the time.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

[HuggingFaceDocBuilderDev]

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.

[fxmarty]

Philipp did not like that kind of dynamicity with perf in mind but I have no opinion, sounds fine to me

[IlyasMoutawwakil]

it only applies when needed so for performance I think there's no added overhead vs what used to be done.
where things can be optimized and i think can be viewed as the optimization oriented path, is i/o binding:

• pre-allocation of output buffers, not during forward but before that, and dynamically changing the size of output buffers when batch size changes.
• decoder models with i/o binding where static cache implementation can be used as output buffers to reduce the overhead of their creation.
• decoder models input/output synchronization which can be before and after the generation loop instead of each forward call (if that's possible).

[fxmarty]

I kind of agree, just not sure what each ORTModelForXXX is for then

[tengomucho]

FWIW, numpy 2.0 has just been released, and int size in windows is now 64 bit! https://numpy.org/devdocs/numpy_2_0_migration_guide.html#windows-default-integer