Ability to Async blocking computations

[blester125]

There are a few places where we can get easy speed boosts. We have a mix of speed via concurrency and speed via parallelism/c-extensions that don't 100 percent mesh. For example, the LSH computation is fast because it was written in Numba but it is used in a blocking fashion so we can't make progress on other async tasks while it is calculating.

Something like an asyncify function which tosses blocking functions into a thread executor (assuming they have big sections where they release the GIL) should let us await on thing like this (or things like the SVD in low-rank updates) and increase through put.

[blester125]

#216 should help profile things. For example, without the asyncify command, all the of the parameters get hashed before any of the later async work can be started. With the asyncify command, we are seeing more interleaving but on the models in the end2end test it did not seem to make a large difference. Will test with larger scale models.