Full text search optimizations

[Ankur Goyal (ankrgyl)]

The goal is to reduce the amount of work (specifically I/O) required to prove that a phrase (which may have many terms of varying frequency) cannot match a set of documents. This basically optimizes for needle-in-a-haystack phrases.

• Look up term metadata before loading positional postings, so missing terms fail fast.
• Probe likely-rarer terms first, with an adaptive per-query ordering based on observed misses.
• Cache duplicate term metadata lookups within a phrase.
• For long exact phrases, run a cheap two-term positional preflight using the lowest-cost term pair.
• Reuse preflight postings when building the final phrase scorer.