Parallelize projection computation and avoid running out of memory

[ethanbb]

I noticed that my CNMF runs were failing due to running out of memory when computing projections. Since I realized the main CNMF run function already has a mechanism for estimating how to break up the movie into chunks for parallel processing to avoid running out of memory, I just reused that here.

Draft because I haven't added tests (should ensure the results are accurate) and probably it would also make sense to process chunks in series in the non-parallel case to still avoid OOM.