Nlists for inference

[chrisiacovella]

Description

This changes how we handle neighbor lists for inference. This includes a more efficient brute force routine and a verlet scheme (although based still on an N^2 building step). These changes will make it easier to support other libraries.

I included a version of the verlet scheme that builds the neighbor list using NNPOps; however I commented it out right now to avoid adding that dependency as it does not seem to provide any improvements in speed/memory utilization after changing around the approach in the pure pytorch versions. I think this can be made faster if we can push more of the required code pieces to the cuda kernel.

box vectors are now stored in the NNPInput class to properly support periodic system.

Todo

• update inference neighbor list tests to point to new routines in neighbors.py.

Status

• Ready to go

[chrisiacovella]

I think this is ready to go. I'm going to hold off on adding back in the multiple cutoffs and do that in a separate PR to keep this smaller and more focused.

[wiederm]

That looks great!

[chrisiacovella]

I added in a basic doc page about the neighbor lists. We will still need to think about the API for controlling things. We might ultimately need a wrapper inside the setup model that allows us to change the neighbor list implementation, instead of it being something that we build the model with (if I built with it, we need to build different versions of the model with different neighbor lists). But we can address that as we work on the openmm interface.