Document what semantic_search results actually contain

[tomchop]

Docs + version bump only; no behaviour change.

Since 2.2.0 the /search/semantic endpoint changed twice (yeti-platform/yeti#1353 and #1354):

• semantic_score is now a bounded 0-1 similarity. It was previously derived on the assumption of a cosine distance when the index actually uses squared L2, which produced negative scores for genuinely good matches.
• Results gained matched_on, naming which of an object's indexed documents matched — self, or approach:N for a DFIQ question's Nth approach, now that questions are indexed as several documents.

Both already reach callers, since semantic_search() returns the API's sections verbatim. The client just never described them.

The range is the part worth documenting: the endpoint deliberately applies no relevance cut-off, because a sensible threshold depends on the corpus and query. Callers are expected to filter on semantic_score themselves — which they can't do safely without knowing what it ranges over. A caller who assumed 0-1 under the old scheme would have discarded most of their real results.

Verified against a live 2.8.0 instance: a DFIQ query returns semantic_score: 0.7714, matched_on: "self".

Bumped to 2.2.1 so the corrected documentation reaches PyPI users.

Test plan

• python -m unittest tests/api.py — 31 tests pass
• No code changed outside the docstring and the version string