Adding HNSW library for fast approximate nearest neighbor search

[smauq]

Hard copied the headers for the HNSW library into maplab. This is a speedup of about 10-30x over the FLANN implementation from OpenCV for floating point descriptors. Summarizing changes:

• Added HNSW with unit tests
• Removed redundant search engines (FLANN, KD-tree, and inverted index)
• Removed unused code in product quantization
• Merged a few packages together to compress the build tree and make this easier to follow
• Re-enabled some old tests that had been disabled for the inverted multi-index
• Reset the min_inlier_ratio flag back to the default of 0.2 as the reduction there was mistakenly done because of the poor performance of FLANN

[ethzasl-jenkins]

Test PASSed.

[ethzasl-jenkins]

Test PASSed.

[ethzasl-jenkins]

Test PASSed.