[WIP] Speed up python calculator

[bernstei]

Avoid needlessly slow operations

• replace list.index() calls for node_types and neigh_types with pure numpy array syntax
• replace pair_forces np.asarray with faster np.fromiter(..., dtype=...) [maybe not faster?]

TODO

• add skin mechanism for recalculating neighbor lists less frequently, as in https://github.com/CheukHinHoJerry/torch-dftd or https://github.com/vicvaleeva/FourierD3/tree/main/src/fourierd3
• other things?

closes #92

[bernstei]

@wcwitt do you think there's enough test coverage that if I were to have broken something and messed up the output numbers it would be detected?

[bernstei]

Main remaining places for plausible speedup are conversion of self.evaluator.node_forces list into an array pair_forces, and calculating neighbor lists less often.

The neighbor lists we can do with skin, but I'm a bit worried about the implementations linked to above: could whatever agent you used to find the issue with the lammps nn list write a good test? The logic that the D3 implementations are using are different enough from lammps that I'm not sure they're getting that right

I'm less sure about node_forces, and I don't really understand why converting 5e5 floats takes an appreciable amount of time anyway. Is it possible to pass that list back as an array directly, through some numpy/ctypes magic?

[bernstei]

I think I've convinced myself that the neighbor list skin heuristics in https://github.com/vicvaleeva/FourierD3, at least with the fixes in vicvaleeva/FourierD3#9, are correct, and could probably be easily incorporated into Symmetrix

[bernstei]

I have a version of the neighbor list skin recalculation heuristic. It's not a dramatic savings, but it's not negligible. With my 2000 atom system, after the speedups in this branch the neighbor lists are about 1/2 the remaining time. The skin heuristics drop the actual neighbor list construction down by a factor of 10 or so, but unfortunately just using the stored info to construct the current neighbor_list output (ijdD) takes almost half the original calculation time, so the savings is about 1/4 of the total time (1/2 of the 1/2 that is neighbor lists). If you think that's worth it, I can clear it and add it to this PR