musdb 0.3.0: a better fit for deep learning datasets

New musdb is released which addresses a few things related to when using musdb inside your deep learning framework.

New features

• musdb.DB() directly loads the list of tracks. No need to call load_musdb_tracks.
• iterating over the track is simplified since the musdb.DB object is iterable and supports indexing
• musdb.tools now has a tool to convert a stems dataset to wav (or flac) automatically.
• we now provide a predefined train/validation split to foster reproducible research.
• musdb now supports chunking of the audio (both via stempeg for stems and soundfile for wavs) to efficiently load only parts of the audio
• A 7 seconds preview version of the musdb18 dataset is automatically downloaded and can be used on the fly. (this is fun for jupyter/colab!).

Incompatible changes

• support for python 2.7 was dropped
• musdb.run was removed since it was not used a lot and people used their own methods to do multiprocessing over different tracks.