Add model training with mini-batches

[JMorado]

This PR adds support for training EMLE models using mini-batches. As training datasets continue to grow in size, this feature becomes essential to avoid memory issues. Unlike with the QM7 dataset, we can no longer fit everything into memory.

The implementation introduces three flags: --use-minibatch, which enables/disables mini-batch training; --batch-size, which specifies the size of each mini-batch, and --shuffle, which shuffles the training data. By default, training still uses the original full-batch optimization.

[lohedges]

Thanks for this. Ignore the test failures. This is because sqm is currently completely broken with recent versions of ambertools. (There are glibc issues, so the package will likely need to be rebuilt.)

[JMorado]

I've added a proof-of-concept implementation to perform the IVM and AEV calculations and to make the training step of valence widths "lazy", i.e. such that batches of masked AEVs are written to disk and loaded on the fly as needed. This is necessary because it is otherwise impossible to load large datasets into memory (the training does not go past the AEV computation, and it's impossible to store the aev_mols tensor in memory). I've been testing this on a dataset with ca. 0.5 M configurations, and it seems like a viable solution so far, although not the most performant. I'm keen to improve the implementation, so any suggestions are welcome!