Improved coordination of CUDA contexts with PyTorch

[peastman]

Fixes #42. This requires the changes from openmm/openmm#3258. We might want to wait a while before merging it, since it's not compatible with OpenMM 7.6.

[raimis]

cuda-memcheck reports several problems. I'll investigate them further. Meanwhile, I leave openmm/openmm#3272 to you, @peastman.

[raimis]

peastman#1 might be a solution, but let me check after rebuilding everything.

[raimis]

The theory behind:

• We have two CUDA contexts and a results there two CUDA streams for each
• Switching between contexts, doesn't synchronise the streams. So an explicit synchronisation is required.

[raimis]

OpenMM 7.7b is out!

How shall we proceed with this? Can conda-forge build the beta of OpenMM-Torch using the OpenMM beta as a dependency?

[peastman]

Sure. I think it should be possible to build a beta that uses the OpenMM beta, though I'm not sure of the details. It would need to use conda-forge/label/openmm-rc as its top priority channel.

[peastman]

Note that CI is failing because it's building against the released OpenMM. I'm going to merge anyway, since it will be necessary for building a beta.