Raise warning for torch double backward

[frostedoyster]

After discussing #139, @Luthaf and I came to the conclusion that we can't do much without sacrificing a lot of performance in the important case where only mixed derivatives are needed. Therefore, we're adding a warning and thoroughly documenting the potential issues that can arise

[github-actions]

Here is a pre-built version of the code in this pull request: wheels.zip, you can install it locally by unzipping wheels.zip and using pip to install the file matching your system

[ceriottm]

Makes sense, we can't get around the limitations of torch.

[Luthaf]

The changes looks good but CI failure is strange here

[ceriottm]

The changes looks good but CI failure is strange here

Guess you'll have to test this - neither Filippo nor I have eaten from the tree of vendor lock-in

[ceriottm]

ok I reverted all my foolish attempts to fix the CI

[Luthaf]

I can not reproduce the issue locally, but I just realized that CI was still using macos-13 runner, which is x86_64. Torch is no longer supporting this architecture, so I switched to macos-14 runner, which is arm64. If this fixes CI I think we are good to go, this could be an architecture-specific bug, or a bug with older versions of torch.

[frostedoyster]

Thanks for the fix!