Add PyTorch 2.5.1 to CI

[BenjaminBossan]

• Drop 2.1.2, update 2.4.0 to 2.4.1
• Drop Python 3.8, add Python 3.12
• Filter new torch warning about weights_only as that trips up some warning tests
• Unpin protobuf from dev requirements
• Filter warnings caused by usage of utcnow by protobuf
• Skip torch.compile tests when torch < 2.4 and Python 3.12 are used, as this is not supported by torch

Note that I had to upgrade torch and Python at the same time because torch 2.5 requires Python > 3.8 and because Python 3.12 requires torch > 2.1 (which is dropped in favor of 2.5).

[BenjaminBossan]

CI needs #1070 first, as torch 2.5 is not available for Python 3.8.

[thomasjpfan]

Assuming two warnings by coming out of this filter can be a bit flaky. Can we directly pull out the warnings from training and validation warnings?

[BenjaminBossan]

Yeah, I thought so too. Do you have an elegant solution for that? I was wondering if there was an API for setting filters to recwarn but I didn't see anything.

At least, with this implementation, we catch unexpected warnings ;-)

[thomasjpfan]

For this specific test, can we filter for the warning we want to check for and make sure there are two of them:

        w_list = [
            warning for warning in recwarn.list
            if msg_substr in str(warning.message)
        ]
        assert len(w_list) == 2

[BenjaminBossan]

After spending too much time fiddling with pytest.mark.filterwarnings without success, I rewrote the filters in a similar way as you suggested to make them more robust.