Add support for training

[wolny]

• add support to embedding networks during prediction (convert embeddings to affinities before segmentation step)
• add dense training support
• test dense training
• support training from headless with --config option
• make sure that trained model can be added as custom model in PlantSeg

Currently only dense training is supported where the user select the dataset directory. It is expected that the DATASET_DIR contains train and val subdirectories, where the training and validation h5 files can be found. In order to support training from exported proofreading, we could e.g. create an additional metadata file inside the DATASET_DIR, which can be parsed for creating the loaders.

The outcome of the training is the model files (config_train.yml, best_checkpoint.pytorch, 'last_checkpoint.pytorch`) saved inside the directory specified by he user.

[wolny]

Hey @lorenzocerrone, I'd say this PR is good to go and we can move forward to combining it with your dataset management functionality in #161.

Quick update: I've removed the training widget from GUI and allow the training in the headless mode only, so that this branch doesn't impact any of the current functionalities and it's safe to merge. Training the model headless and then using it in napari gui works already, but would be great to have it properly integrated with the dataset management :)

[lorenzocerrone]

Looks good! Let’s have a minor release, and plan a major one by the end of January when the dataset manager and the docs are ready :)