Integrate with pytorch-lightning

[mpariente]

Solves #3 #6 #16 .

• Adds interface with pytorch-lightning through asteroid.engine.system.System. Automatize the interface while keeping core objects (model, optimizer, loss and data) separated. Keep everything overwritable.
• New recommended way to instantiate model : make a function which will return a class. The user write the class inside the recipe. Enables easy saving/loading for evaluation or re-training while keeping class parameters clean and transparent.
• Removed Container (trying to cover all forwards is not feasible), Solver (lightning will do training better than us) and SubModule (saving and loading are solved by the above).
• Rewrite ConvTasnet and FilterbankDesign recipes accordingly.

About pytorch-lightning:

• Training is done seamlessly on CPU, GPU or set of GPUs.
• Resuming is done by default (also practical with the new save/load method)
• Early stopping + Model checkpoint + Loggers (dig deeper) + Tensorboard + learning rate schedulers

Low priority : dig deeper in the features to give sensible default values to asteroid users.