LLPR improvements

[frostedoyster]

Implements a set of improvements to the LLPR:

• Zero-cost algorithm to include forces and stresses in the covariance matrix
• Force uncertainty example
• Dtype handling
• Torchscript test

---

📚 Documentation preview 📚: https://metatrain--307.org.readthedocs.build/en/307/

[SanggyuChong]

Please see comments about how the parameters are supposed to be interpreted, defined, extracted, and used.

[SanggyuChong]

see later comment

[SanggyuChong]

The logic behind the introduction of parameters is very unclear. What are the parameters? How does the relate to the last-layer features? How are these parameters to be defined, and then extracted for the covariance matrix to be correctly calculated? This needs to be documented.

[SanggyuChong]

To me it looks good, didn't understand the shift from computing something on the energies and forces to working directly on the loss functions, but now i understand and see it now.

Thanks for the revamp, filippo, looking forward to playing with this!