Improve build from source instructions

[sef43]

Address #64

• Provides a conda environment file
• update build instructions in README

[raimis]

Please, don't use stuff from pytorch.org. Their binaries are incompatible with conda-forge.

[peastman]

We probably want to reconsider the build process and possibly make corresponding changes to the CMake script. If you set TORCH_DIR (not PYTORCH_DIR) to the directory containing TorchConfig.cmake, it should be able to automatically find everything it needs.

[raimis]

Just reuse the one from CI: https://github.com/openmm/openmm-torch/blob/master/devtools/conda-envs/build-ubuntu-18.04.yml

[peastman]

CUDA is only needed if you have a NVIDIA GPU you want to use with it.

[peastman]

These instructions seem to be Linux specific. We also support Mac and Windows. Note that conda-forge does not currently provide PyTorch builds for WIndows, so in that case it's necessary to get it from the PyTorch website or the pytorch conda channel.

[sef43]

I have made specific instructions for Linux (with and without CUDA) and MacOS.
One issue is that I think the conda-forge NNPOps only has linux with cuda versions available

[raimis]

We should put the from-source installation instructions to a separate file (e.g. INSTALL.md) and link it from the main READM.md. Most users can install with conda, so they don't need to be distracted.

[peastman]

One issue is that I think the conda-forge NNPOps only has linux with cuda versions available

That's because conda-forge doesn't provide PyTorch builds for Windows. That makes building from source especially important, since it's the only option for Windows.

[sef43]

For instructions for a building on Windows we will need to update PythonInstall and python/setup.py etc to work with windows. I am not familiar enough with the build system/Windows to do this.

I have been able to compile the c++ code on Windows (CPU only) using visual studio and this environment:

name: build
channels:
  - conda-forge
  - pytorch
dependencies:
  - openmm >=7.7
  - pip
  - pytest
  - python
  - pytorch::pytorch
  - pytorch::cpuonly 
  - swig <4.1

One quirk is that pytorch requires MKL, but the conda mkl for me did not seem to have all the files needed to compile so I had to download the full MKL installer from intel.