Before finishing this PR, we should discuss the default behaviour of the beam importers that import numpy data. Right now there are actually in an inconsistent state:

• When from_ocelot is used with default dtype=None, it uses the dtype of the numpy array passed to it. In my opinion this is the correct behaviour and it mimickes the dtype one obtains from torch.tensor when passing a numpy array. In fact, the behaviour is implemented using torch.as_tensor to determine the appropriate dtype.
• from_astra on the other hand uses a default of dtype=torch.float32. This does not follow the automatic dtype selection we are implementing in this PR and thus definitely has to change. I see two possibilities:
  1. Simply change to dtype=None, making its behaviour identical to from_ocelot. This would be my preferred choice but it leads to a number of test failures currently. The problem is that the tests implicitly assume a beam of type torch.float32 and the solution would be to explicitly set dtype=torch.float32 in the tests while loading the beam.
  2. Change to dtype=None, but query torch.get_default_dtype in that case to determine the dtype. This makes the API mostly consistent with the current behaviour since torch.get_default_dtype() defaults to torch.float32 but deviates from how torch.tensor determines its dtype from numpy arrays.

In case we go for the second option, from_ocelot should be updated in the same way.