Add ability to limit ITQ training memory usage

[predicative]

This is #370 again, but this time against master.

Copying from there:

This PR adds the ability to limit memory usage [during] ITQ training. The current implementation modifies the train_itq.py tool to have an option max_descriptors that, when set and if necessary, sub-samples the available descriptors to the desired amount.

... [snip] ...

I'll also note that there may be some other opportunities to save memory (namely by del'ing descriptors in ItqFunctor.fit after its last use; see below), but this seems like good functionality to have anyway.

SMQTK/python/smqtk/algorithms/nn_index/lsh/functors/itq.py

Line 330 in ea94e3b

x = elements_to_matrix(descriptors, report_interval=dbg_report_interval,

I've combined the commits from #370 into one and added a release note.

@Purg, would you be interested in seeing tests for this, and if so, could you advise on how best to do so? My initial idea would involve a rather large amount of mocking.