By-receptive-field interpolation at predict time ; parameterization of transforms ; PointNet++ support

[CharlesGaydon]

Interpolation first happens only within a receptive field ; interpolated logits are then summed if multiple predictions were made for a single point.
Evaluation of IoU is simplified in test to only be computed on predicted points (that can be synthesis points from gridsampling).
Transforms can now be configurated from config files, to e.g. add augmentations.
PointNet++ is supported, and is tested with an overfit test on the toy dataset. This is to ilustrate that models from https://github.com/pyg-team/pytorch_geometric/tree/master/examples can easily be integrated. One must be careful for model parameters, and how they interpolay with data transformations (e.g. adapting radius if the positions were normalized).