Implemented new Difficulty Predictor model that uses ONNX Runtime instead of PMML to enable LDA in the model.

[Brollyy]

109

The new model has way better predictive power on lower diffs and is slightly more consistent with 4-6 diffs.

I removed the "Treat maps as WIP" checkbox, since with the new metrics, the model always takes the time between first and last note, so there was no difference.

There were some issues with very hard maps, where the predicted difficulty was very low, so we've opted in to display a warning instead when the map parameters go outside of range on which the model was trained:

EDIT: after remodelling due to some implementation issues before, the new model is slightly improved. It uses these features:

1. Note density of the whole song.
2. Average time differences between notes.
3. Number of time differences below 2 seconds (essentially counts how many notes are "close" to each other to account for longer breaks).
4. 95th percentile of the local note densities in a rolling 4 second window (2 seconds before note, 2 seconds after note).
5. 40th percentile of time differences between notes.