Reduce memory consumption of PAV backend

[osma]

Currently the PAV backend collects the subject vectors for each document into memory. This takes up a lot of RAM and limits the number of training documents it can handle. Several improvements could be made in this area:

• limit the precision of subject vectors (float32 instead of float64)
• use sparse arrays for the subject vectors
• when collecting the subject vectors training the PAV backend, store them on disk (e.g. in a temporary file or LMDB database) instead of in RAM

[osma]

The float32 part was implemented in #340; other ideas have been split up into individual issues. Closing this one.