Thingsvision

- added `output_type` option for choosing between `torch.Tensor` and `np.ndarray` for type of returned features in `extractor.extract_features(...)` for substantial speed-ups during the extraction process (moving tensors to CPU is costly which is why we want to avoid it wherever possible)
- note that the above is only possible for `backend="pt"`

- memory issue fixes
- abstract base class
- removed mixin classes

- incorporated `torchtyping` for better data type and shape annotations

- fix for custom model

- file name fix for custom model

- new custom model