More granular type hints for numpy.ndarray's

[ascillitoe]

There are numerous places in alibi-detect where we type hint method arguments with the rather generic np.ndarray, but actually only expect certain dtypes. For example the upcoming FET detectors, and SpectralResidual (see #382). This could be dealt with purely through a little more documentation, but it might also be advantageous to explore more granular type hints.

The above is related to SeldonIO/alibi#513 and SeldonIO/alibi#511, or we could use numpy's inbuilt capability (https://numpy.org/doc/1.21/reference/typing.html#numpy.typing.NDArray for dtypes, and at some point numpy/numpy#16544 for shapes). The latter would however cause issues with old numpy version compatibility.

[jklaise]

Yea unfortunately we can't just update to numpy>=1.21 and there also doesn't seem to be a backport of these new numpy types. I think it might be a good use case of beartype validators.