Sparse fast alpha and gamma informational distances

PHATE now implements a sparse, fast alpha decay kernel which has minimal memory requirements, compared to the old alpha decay which required a parwise distance matrix. Alpha decay now runs by default. In order to use the k nearest neighbors kernel, run with a=None.

Other notable changes:

• graphtools implementation of kernel matrices
• gamma replaces potential_method for the selection of informational distances. gamma=1 is equivalent to a log potential (default) and gamma=0 is equivalent to a square root potential
• arguments alpha_decay and potential_method are deprecated.
• PHATE is now implemented as a visualization method in Scanpy: https://scanpy.readthedocs.io/