[pytorch] Add encoders for X

[ebezzi]

Description

It would be useful to add custom encoders not only for obs, but also for the X matrix. This would allow e.g. to let the user decide whether they want a sparse or a dense without relying on a flag (currently return_sparse_X), provide a different formats for sparse matrices, etc.

[pablo-gar]

We should have the encoder be flexible to allow on-the-fly tokenization of data. Form @AlejandroTL we heard the following:

Only thing I am missing [in the PyTorch loaders] a priori is something to handle different tokenization strategies during the dataloading. For instance, retrieving the counts, retrieving the indices of the genes given some previously stored dictionary, or binning the data on the fly, or creating a ranked list, etc.