Local Ancestry Initial Implementation (bystrogenomics#396)

This implements a deep learning version of RF mix with substantially
better scaling properties. In this case, most of the computation time
scales independently of the number of segments considered. The model
functions by training a neural network for each region of the genome,
but all the layers after the first are completely shared. This means
that we learn stochastically both by sample and by number of regions
dramatically increasing computation speed.

---------

Co-authored-by: Alex V. Kotlar <[email protected]>