Add a heat flux wrapper

[GardevoirX]

Closes #104

This is still in a very early stage and subjected to change.

TODO:

• Interface for getting the heat flux from the model
• Checks for the applicability of unfolding the system
• Tests
• Documentations
• Remove Replace the dependency on vesin.metatomic with vesin.torch and rebuild neighbor list from the un-unfolded system (Hard to build the neighboring information between the replica atoms, e.g. atoms in different image boxes. And vesin.metatomic is not supposed to be used in metatomic.)

Contributor (creator of pull-request) checklist

• Tests updated (for new features and bugfixes)?
• Documentation updated (for new features)?
• Issue referenced (for PRs that solve an issue)?

Reviewer checklist

• CHANGELOG updated with public API or any other important changes?

[GardevoirX]

I think this is now ready for reviewing. Though the current results with pet models are not good enough, soap-bpnn still gives some promising results. Plus, there's no evident bug in the current implementation. Perhaps a thorough checking for the code can reveal underlying problems.

In terms of the implementation, thie current state hasn't reached an ideal state. The calculation of the term1 of the heat flux can be implemented in a JVP style, but it is still in VJP and calculated through a for-loop. This might be improved after metatensor/metatensor#1063 is merged, but previous tests suggested that more blocks are in the path.

The second thing is how we do "unfolding" in _unfold_system. There could be two ways:

1. The current way: check if particles are close to the box boundary, and generate the replicas if true.
2. Do a NL calculation to get the atom pairs closer than the interaction range of the model, and then only unfolding the atoms in the pairs with non-zero shifting vector.

Tests showed that the second way can generate fewer atoms and thus can save runtime from the model inference, while an extra NL calculation also costs time. My intuition is that which one is faster depends on the system size.

Also maybe we need a tutorial about how to use this wrapper.