Improvement of the difficulty predictor

[Brollyy]

#120

• Updated Nytilde's model to the newest version and added the fallback model.
• Slight refactor to prepare for implementing different difficulty predictor algorithms.

Summary

The new model doesn't seem to improve too much from the earlier one - the predictions for typical maps around lower and mid levels seem ok (leaning towards underestimation while I looked at some random maps), but for the high difficulty maps it goes off the rails, even with the fallback model:

Ira Sancti (verified manually that the main model is called for all diffs)

Through the Fire and Flames (verified manually that the fallback model is called for all diffs)

Parallel Universe Shifter (verified manually that the fallback model is called)

TODO

• [ ] The new fallback model still has the same issue as the full model - for maps outside of the training range for features, the difficulty prediction is decreasing with the new notes being added.
• [ ] Re-implement PKBeam's ML model and implement option to select the model
• [ ] Check Melchior's standalone algorithm