also log a warning if GLM.fit() doesn't converge

messages now identify the params that caused assertions to fail

some assertions are only expected to pass for reg_lambda=0.0
where the true model can be reconstructed almost perfectly

Needed to increase the required loss trace improvement threshold
from 1e-7 to 1e-3 to deal with occasional noise in cdfast steps

beta is updated by GLM._prox, which invalidates the cached z

the only place that uses _cdfast immediately invalidates the cache

to match package style

Disabled flake8 checks:
  W504 conflicts with W503-type style used in the package
  W605 conflicts with latex code in Sphinx docstrings

ensure _logL calls use z argument by name

use scipy.special.comb instead of removed scipy.misc.comb