Dey+22 Calibrated Predictive distribution

[torluca]

The goal is to implement the calibrated predictive distributions of Dey+22 in rail.evaluate

[eacharles]

Quick note to consider if we should split this in to it's own package b/c of pytorch dependencies

[aimalz]

Thanks for sharing your great progress @torluca! In response to today's discussion, I anticipate a question to resolve as you stage-ify the code: while the calculation of the PIT on subsamples is like evaluating a metric, I think the calibration process that outputs a catalog of photo-z PDFs is actually more at home as an estimator, so it might be appropriate to move it to the other subpackage at the same time.