Use distance layers of HEALPix map, rather than volume rendering, as model input

The model currently uses the 2D plate carée false-color image and the volume rendering as an inputs. Both are computationally intensive. The latter is very computationally intensive and OpenMP parallelized to many threads, and also throws away a lot of information because it is a projection.

Investigate using the HEALPix map itself --- flattened to a modest fixed resolution --- as model input instead. Or, if the topology of the neural network requires the data to be on a 2D plane, then use the 2D plate carée images of the probability and distance layers as input.