Upgrade PZFlowPDF, the pzflow-based RAIL Estimator

@sschmidt23 implemented a basic pzflow-based RAIL Estimator in PR LSSTDESC/RAIL#116 (addressing issue LSSTDESC/RAIL#94). There are numerous upgrades that could be made. I will list them here to keep track of them:

[ ] fix slight jitter in redshift mode values between python 3.7 and 3.8
[ ] handle non-detects and non-observations differently
[ ] use another normalizing flow to marginalize over missing magnitude when training
[ ] use an error model that is Gaussian in flux space rather than magnitude space
[ ] change the flow into a conditional flow for increased speed

This list is not necessarily exhaustive, and I will try to update the list as we think of more things.

LSSTDESC / rail_pzflow

Upgrade PZFlowPDF, the pzflow-based RAIL Estimator #4