systematic treatment of floating point warnings from numpy

[sergpolly]

it would be worth considering a systematic treatment of floating-point related warnings - division by zero and invalid operation pollute output too much - esp whenever obs/exp is involved - classification of float-point errors https://docs.oracle.com/cd/E19957-01/806-3568/ncg_handle.html

1. consider using (on a per module basis ... or in __init__ or wherever)

# set new err-stated, to hide division by zero warnings
_err_state = np.seterr(divide="ignore", over="raise", under="warn", invalid="ignore")

2. or use a conext manager https://numpy.org/devdocs/reference/generated/numpy.errstate.html for every obs/exp (and other relevant operations) - not sure if there is a big performance penalty associated with it ...

with np.errstate(divide="ignore", invalid="ignore"):
    snippet / e

3. or use np.dividehttps://numpy.org/doc/stable/reference/generated/numpy.divide.html with the where filter engaged - to pre-emptively fill those anything/0.0 with np.nan