V2.4.0 - refactor and hdf5 datamodule

[CharlesGaydon]

Key Changes :

• Validation (test/val) IoUs are computed on the interpolated sample instead of the subsampled one, which takes longer but gives more meaningful information during training. This means that preparation transforms are defined separately for training, eval (val/test), and prediction.
• HDF5 datasets for training/evaluation, and Inference dataset that uses part of the code. Overall simplification of data preparation. It now happens on the fly, and input is a folder of LAS + a csv file with basename (xyz.las) x split pairs (split=train/val/test).
• Only keep RandLaNet to ease maintenance, as other architectures are not used.
• Only keep FrenchLidarHD points_pre_transforms, as Swiss data is not used.
• Get rid of unnecessary, sometimes long tests, that are a burden to maintain (overfitting test in particular).

All changes are retrocompatible with previous models.