Add dtype parameter to External

[RaulPPelaez]

This allows to use External with any dtype.
In addition to being useful for TorchMD integration, I normally use External to put models into CUDA graphs, as it handles all the nuisances transparently.

[RaulPPelaez]

@AntonioMirarchi do you see any issues with this? It should be retrocompatible.

[AntonioMirarchi]

From a users point of view maybe we should consider to use "single" and "double" precision as init args and then a dict to convert them. I use the external module from a yaml file in torchmd or openMM and I cannot specify a torch dtype.

Furthermore, usually platform as torchmd and openMM have their own control over the precision so I don't know if maybe some assertion is needed in order to have compatibility between model and platform (but of course this is out of this PR).

[RaulPPelaez]

With OpenMM one should not use External. TorchForce and TorchMD_Net get along well.

I made it so that dtype can be passed as a string. So in a yaml you can put:

External:
   dtype: float32

Or any valid torch.dtype (double, float64, half,...)

[AntonioMirarchi]

Niceee, I will try to run something to be sure but seems fine!

[AntonioMirarchi]

LGTM, I can run sims with it using torchmd