WIP

This will likely only work with a CUDA compatible GPU. While OpenMM has support CPU, it chooses different code paths. See for example openmm/openmm#3288 where the poster runs into different errors depending on whether the platform is CPU or GPU.

Notes for myself

• yank does not linearly scale the reciprocal contribution from PME. Instead, they reweight the endstates by increasing the cutoff a lot. See this function and the places it's called for info on how to use it. To a maximum of 16Å