IVFSparse Improvements

[davidmezzetti]

Update IVFSparse with the following improvements for speed and accuracy.

• Build k-means clusters using top n most important features from L2 norm vector
• Prune small clusters and merge them into larger clusters
• Calculate block summary vectors using max values per column (block max summary vector)
• Add threaded batch search (majority of code is executed outside of the GIL with SciPy/NumPy, someday won't matter with GIL-less Python)