Average match score across talent

Mean match score across the same talent GET /talents/count would count, with the same filter/vectors/proximity/audience parameters and the same visibility rules. Split off the count (POOL-2184) because averaging needs a per-document value and so can never be answered from the index, which made every talent-table tile render read the whole tenant. Only the vector-scoring stages write score, so a query without vectors returns null — including a proximity-only query, since proximity compiles to a bare $geoWithin $match that filters but writes no score. It also returns null when the V3 engine is primary for the target opening: the number this route produces is a V2 average over the matched population, while the rows of GET /talents come back V3-scored, so the two would describe different engines. Average the score on those rows instead. On such a tenant this route therefore serves no number at all, and the same null comes back from the deprecated averageScore field on GET /talents/count — the two agree by construction, including on refusing a filter no index can serve.