Update, extend, cleanup distrcompositor

[RaphaelS1]

Following changes are required:

1. Remove composition from crank to distr - This doesn't make any sense for abstract rankings, composition can only make sense for lpto distr.
2. Add composition from response to distr - This can be most efficiently done by abstracting the probabilistic regression composition and using the same functions in both

(1) is higher priority as its results are meaningless.

[fkiraly]

Remove composition from crank to distr

Why? I can imagine cases where this makes sense!

The most obvious case would be calibration methods, e.g.: stick your predicted rank into a distribution as shape and/or location parameter; then do grid-tuning (or gradient descent tuning if/once autodiff is supported).

This is very common for probabilistic classifiers, and the probabilistic regression/survival counterpart is also not absurd.

Here's probability calibration in sklearn: https://scikit-learn.org/stable/modules/calibration.html

[RaphaelS1]

Sorry I should clarify this:

The current composition assumes crank = lp and then uses a semi-parametric composition, e.g. h(t) = h_0(t)lp for baseline h_0.

However the new composition type to end (point 2.) instead uses the composition you describe, which can handle crank

[fkiraly]

ah, makes sense.