Object-based, multiprocess, and caching pdet and weight estimator

[kcroker]

Implements the core functionality in pdets_from_grid() as a persistent-state object. It will cache the underlying trained regression model, so that training the model only needs to happen on the first call. Subsequent evaluation of the model can also be distributed across a multiprocessing.Pool. Multiprocessing can be memory intensive, because every child needs a copy of the trained regressor, and that's ~350MB. Because the regressor is complicated, getting the memory consumption down will probably require something like map_coordinates() with a shared_memory numpy array. If this approximation scheme delivers the necessarily accuracy, it would likely be 10-100x faster in addition to consuming 1/cores less memory.

Usage example:

from selection_effects.utils.predict_detection_probabilities import LVKWeighter

# Initialize the object with "filename:key" string
# Defaults to loading a cached version if present
lvk = LVKWeighter("pdets_grid.hdf5:O3actual_H1L1V1")

# Estimate with some data
with multiprocessing.Pool(10) as p:
  data['pdet'], data['weight'] = lvk.estimate(data, pool=p)

[kcroker]

Apparently multiprocessing.Pool is not the preferred way to parallel out scikit learn things. It may run much faster if implemented with joblib. See https://scikit-learn.org/stable/computing/parallelism.html Also claims that it shares memory too. Not sure if this applies just to training, or also to prediction...