cleaning + refactoring to get training scripts working with dumped data

refactoring to align config.json structure, dumped data format and training scripts.

• move paths spread on several json into one central config.json
• modify slightly the keys from training scripts cfg files to be able to merge json files into a unique config. The final config is dumped when training starts
• modify the merge of json files in main.py (use jsonmerge instead because was not working on my machine :()
• modify the dataset.py to align with dataset dump format: a single tensor (chroma is already upscaled etc.). Currently not taking advantage of these data preparation to speed up cpu processing
• adapt loss and model to take into account chroma sizes
• rename bpm into ibp to comply with format
• put quantizer for each components inside the model.json parameters
• slightly modify the directory structure:
  • traininig: main.py, dataset.py, trainer.py, logger.py, training_default.json, stage1.json
  • model: model.py (was net.py), model.json
  • config.json
• command line adaptation: see readme.md
• removed multi dataset/validatation set for now
• pass pep8

training is testable by following the steps in readme.md to create a mini dataset and use it for training (no sanity check done yet)