Enable torch

[dobraczka]

Since Faiss has torch support with tensors that are on the GPU, it is desirable to avoid moving these to and from the GPU.
This PR adds the ability to have an entire pipeline for hubness reduced nearest neighbor search on the GPU.
For some hubness reduced methods it is not feasible to do this (probably only MP-empiric).

TODO:

• Adapt TypeHints
• Try DisSimLocal adaptation

[dobraczka]

Since it seems to be more or less impossible to provide type hints for a library (torch), that may or may not be available at runtime, I've given up and used T = TypeVar("T") to at least show that specific output tuples have the same types. I did not do this for inputs, since mypy will complain that T does not have this or that function or that a specific numpy function expects an ArrayLike etc. There may be a way to get around this with instance checks and asserts, but at the moment it is too much effort with little reward.

On the upside: All HubnessReduction classes now can use torch tensors. Some more cleverly than others. All NNAlgorithm classes except Faiss will complain if they do not get numpy arrays.