QoL improvements

[lgleznah]

This PR addresses multiple issues I have found during my own experiments with this repo. Hopefully, these will be helpful for those who are working with this technology, which so far, looks promising for doing point cloud stuff.

This is a quick summary of all the stuff I have added:

• Validation max epochs is now interpreted as stated.
• Checkpoints are now saved of the current best model in terms of mean IoU.
• Option for specifying class weights by setting cfg.data.compute_class_weights = True. Addresses Imbalance datasets tackling #9
• Training and validation losses visualization for monitoring potential overfitting during training.
• Validation step now shows confusion matrix.
• Testing script now generates a PLY file with the predictions, for visualization purposes.
• Other minor bugfixes.