Wordbatch

Queue Ray backend, so larger numbers of tasks can be sent to Ray at once

32-bit precision version of FTRL, and model finalization for 1/3 memory footprint

Enables use of Apply as a decorator with simplified API. Adds caching and basic Numba support as options for accelerating Apply minibatch processing

Several minor fixes and updates

Support for several distributed backends: Ray, Dask, Spark. Major refactoring and API change.