single vs. double precision

[mastoffel]

in _check_input we're currently converting to float32 due to the PyTorch models. Maybe better to use double precision where possible and single precision only where necessary, e.g. for deep learning models / models using GPU's. Thanks @edaub.