Parallelize Everything

[alejandroschuler]

I think it would be useful to parallelize all of the code, since parallelized code can still be run on a single processor. We can get rid of glrm altogether, or keep it for pedagogical purposes. For massive datasets, it currently takes forever to run stuff like observations() and we can turn df2array into df2sharedarray in one shot.

@time data = readtable("data/sample_100000");
elapsed time: 29.124417016 seconds (5721166992 bytes allocated, 7.27% gc time)

julia> @time obs = observations(data);
 elapsed time: 547.430868981 seconds (12869023916 bytes allocated, 93.97% gc time)

[madeleineudell]

Specializing on the new shareGLRM type, which is a subtype of the AbstractGLRM type, should make it easier and cleaner to parallelize larger swaths of the codebase.