Multiprobe L2 LSH first pass

[alexklibisz]

First pass at multiprobe L2 LSH based on paper: Multi-Probe LSH: Efficient Indexing for High-Dimensional Similarity Search, Qin, et. al.
This uses the naive method for picking perturbation vectors. Namely it enumerates all of the 3^k possible perturbation vectors and uses a heap to pick the best ones using the x_i(\delta) scores from section 4.3 in the paper.
Next I'll implement the optimizations from sections 4.4 and 4.5.

This already shows a meaningful improvement in the Sift dataset benchmarks, specifically:

• Mapping with L=300, k=2, r=1, query with 4000 candidates produces p90 recall/duration = (0.73, 54ms)
• Mapping with L=100, k=2, r=1, query with 4000 candidates and 3 probes produces p90 recall/duration = (0.78, 54ms)

So you can store 3x fewer hashes and get better recall with the exact same latency.

This required some internal changes to the way the HashingFunctions are cached.
I also simplified how the MatchHashesAndScoreQuery handles cases with all-zero counts by just returning a DocIdSetIterator.empty().