Add MLIPOps backends

[CompRhys]

https://github.com/openmm/mlipops

[CompRhys]

@yantar92 this might be a better solution to your alchemiops on AMD trouble if this works on AMD.

[lil-lon]

@CompRhys
I'm interested in working on this. I've contributed to the neighbor list code here before, so I started by checking whether MLIPOps' NeighborList can be wrapped as a torch-sim backend.

I found a few things to watch out for, but the important ones are:

• It only considers the nearest image of each pair (minimum image convention), so every box dimension must be at least twice the cutoff. For smaller cells, neighbors through other images are silently missed.
• PBC is all-or-nothing and the box must be in reduced lower-triangular form, so slabs and general cells are not supported.

These box requirements are documented in the MLIPOps user guide, so this is by design rather than a bug, and it can't be integrated as is. I could open an issue upstream to ask whether multi-image support is in scope. Does that sound like a reasonable next step, or would you prefer a different direction, e.g. adding it as an opt-in backend restricted to the supported cases?

[CompRhys]

I think that making the issue upstream would be good, I think that for most the openMM use cases MIC is enough but potentially they will be happy to extend if the usecase of structure optimization is highlighted.

I would also want to strip the alchemiops in electrostatics d3 etc as there's the vendor lock-in of alchemiops there and this package is supposedly faster and not locked to CUDA.