Build variants to be dinamically loaded depending on the platforms available

[pabloferz]

We have support for CPU, Reference and CUDA platforms in OpenMM, but we try to link unconditionally to the CUDA libraries. This PR enables building both a CUDA-enabled and CPU-only variants that can be loaded dynamically on the python side.

OpenMM handles this scenarios by dynamically loading platform-dependent libraries/plugins on the C++ side. We might want to look into considering that in the future.

Other relevant changes from this PR:

• Contains an alternative to remove direct openmm import #23
• Sets the __version__ attribute of the python package with setuptools_scm
• Deprecates loading the package as openmm_dlext in favor of loading it as openmm.dlext

[pabloferz]

Closes #28