Find a better way to implement range queries

[richardstartin]

After "freezing" the classifier, match queries on numeric ranges take between 1-100 microseconds for a range of numbers of segments and feature space dimensions. This comes at the cost of potentially enormous memory usage, since the bits corresponding to all entailed thresholds are duplicated onto the threshold. One alternative is to compute the entailment on the fly to reduce memory requirements, but this reduces query performance at least 10x, and gets worse with the number of thresholds.