Search index is optional for precomputed_knn

[jlmelville]

This is a fix for #848 in the sense that it allows passing only indices and distances to as the precomputed_knn parameter (i.e. no pynndescent object required). Things to note:

• You will get a warning that you can't transform new data if you do this.
• If you then try to transform new data via the transform method, a NotImplementedError will be raised.
• If you have small data and force_approximation_algorithm=False you will not see the new warning, because the knn will not be used. Setting force_approximation_algorithm=True avoids this situation so this new feature works for big and small data.
• docstring for precomputed_knn parameter on the UMAP constructor has been updated.

Tests successes/warns/failures were the same before and after this addition (there is one failure currently which was failing on upstream-master for me before I added any code). I have added a new test to exercise these various cases.

[lmcinnes]

Thanks, this looks like a nice simple way to resolve the issue.