Optimize nested term aggregations with a shared term ordinal cache

[fulmicoton]

When a string term aggregation is nested inside another bucket
aggregation, each parent bucket previously resolved term ordinals
to strings independently via sorted_ords_to_term_cb. This was
redundant since sibling buckets share the same term dictionary.

Introduce TermOrdToStrCache which resolves all observed term
ordinals once across all parent buckets, then stores the results
in a StringArena (a single contiguous String buffer) with
compact StringRef handles. The cache uses either a dense
Vec<Option<StringRef>> or a sparse FxHashMap<u64, StringRef>
depending on the ordinal range.

Also add collect_term_ords to the TermAggregationMap trait so
each map variant can export its live ordinals into a shared set.

Closes #2892

[fulmicoton]

This actually causes a regression, probably because the term ords do not get filtered out via cut_off_buckets anymore, including for the first level of aggregation.