EMLE training

[kzinovjev]

This implements EMLETrainer class and emle-trainer script that exposes training as CLI. The reference data is expected in the same format as emle-analyse (a tarball with ORCA and horton outputs). This should facilitate active learning, where error analysis and retraining is performed for the same data.

Main changes to the common codebase:

• All EMLE-specifie AEV logic is moved to EMLEAEVComputer class. This includes masking, normalization, and mapping between EMLE and AEVComputer zid values (the latter only needed when using a common computer with ani2x)
• A small increment (1e-16) was added in various places in EMLEBase to avoid nan in the gradients w.r.t. EMLE parameters. Before it was not needed, since the gradients were only calculated w.r.t. atomic positions
• aev_mean feature was removed

[kzinovjev]

Ready to go. Changes:

• Numerous bugfixes in both the core and emle-analyze and emle-train scripts and API. Tested training for both QM7-based model and a custom one. Analysis also now works well (fixed multiple unit-related bugs).
• Fixed units of the embedding forces in EMLECalculator. Propagation now seems to be stable.
• Updated docstrings for EMLEAnalyzer
• Added a short README section for emle-analyze and emle-train
• Changed logging to loguru

@JMorado, feel free to add to the list if I forgot anything substantial!

[lohedges]

Many thanks for this. I've merged into my feature_train branch, blackened, and am now running the CI against the updated AEV model. Everything passes locally and the sire-emle interface works without issue.

[JMorado]

Many thanks to both of you. Looks good to me.