feat: generate embedding from index and initialization

[wilsonjr]

This PR aims to reduce memory usage when an index is pre built for the input data. The user is able to generate an embedding given the precomputed index and initialization.

• Generates embeddings using precomputed_knn and init (np.array);
• Refactors fuzzy_simplicial_set by pulling out nearest_neighbor call (so we don't require passing the input data);
• Usage example

[lmcinnes]

This looks quite useful. There are a number of changes, so I'll have to take some time shortly to go through it and verify everything.