Speed up crystal RDF analysis

[InfluenceFunctional]

Currently as or more expensive than running the model itself.

Particularly slow is the atom-to-atom indexing.

[InfluenceFunctional]

Made some progress on indexing in element mode, but still not great.

[InfluenceFunctional]

Got about 30% from fully parallelizing the RDF call (only in elementwise so far)

[InfluenceFunctional]

possibly use scatter or one_hot to substitute the histc - possibly then more flexibility around dists indexing

[InfluenceFunctional]

Got another significant speedup using new scatter histogram method, but it requires a large number of bins (~2k) to converge to old method. Still has one slow step which holds it up by about 10x.

[InfluenceFunctional]

Got about 30% from fully parallelizing the RDF call (only in elementwise so far)

now implemented in atomwise mode

[InfluenceFunctional]

We then also have the issue that computing distances between these for large numbers if crystals is rate-limiting in CSP work.
Perhaps there is a way to do this directly from the edge information, without going through the explicit construction of the RDF?

Certainly, in 'atomwise' mode, the edge graph if fixed, and we could take explicit edge-wise differences.