Change the architecture of salento to one that is feasible to streaming

[cogumbreiro]

Salento expects as an input a sequence of packages.
The problem is that the file format that contains the sequence of packages is a JSON objects, which means that all packages must fit into memory to read them. We currently have some use cases where the datasets do not fit memory, so this architecture is a bottleneck for scalability.

We need to:

1. change the file format to something amenable to streaming packages
2. change the internals (say, train.py) such that data is loaded lazily and use as much as possible generators (versus creating lists upfront)