Export to BMZ and other changes

[jdeschamps]

Change of behaviour

• Mean and std are no longer calculated during prediction.
• Prediction can now be performed without tiling, default behaviour.
• Prevent extra parameters in TransformModels, which was previously causing wrongly passed parameters to not raise an error.

Configuration

• ArchitectureModel to simplify models (UNet, VAE, Custom), now export to dict without architecture.

Model I/O

• HyperParametersCallback to save the configuration with mean and std in the checkpoints.

Others

• TileInformation object is passed along the tiled prediction, replacement to tuple of arrays.
• Normalization and denormalization are now compatible with BMZ (using a eps=1e-6 parameter), and are now Albumentations classes.

⚠️ Addded failing tests for depth=3 in UNet and channel > 1. See test_lightning_module.py.