feat: search the catalog by meaning

[jaspermayone]

Part of #453. Stacked on #639.

q matches the literal words. semantic=true ranks by meaning instead, so
"intro to programming" reaches "Computer Science I". It works on sections and
on instructors, on REST and on GraphQL.

curl "…/api/v1/catalog/sections?q=learn+to+program&semantic=true&term_uid=202710"

{ sections(filter: { q: "learn to program", semantic: true }, first: 10) { nodes { crn title } } }

How a request runs

flowchart TD
    A["q + semantic=true"] --> B{Flag on and<br/>OPENAI_API_KEY set?}
    B -->|no| K[Keyword search]
    B -->|yes| C{Query in the cache?}
    C -->|yes| E[Rank the filtered rows<br/>by cosine distance]
    C -->|no| D[Embed the query]
    D -->|API failed| K
    D -->|vector| E
    E --> F[One page, nearest first]

Loading

Points worth reviewing:

• The other filters run before the vector cut, so a term filter never
  leaves the page short.
• The query vector is cached by the text itself, lower-cased. Students ask the
  same few things during registration week.
• Every failure path falls back to the keyword search. A semantic request never
  returns an error because the API was down.
• semantic_search is a global Flipper flag. Turning it off stops every
  embedding call at once.
• A new Rack::Attack throttle limits semantic REST requests to 30/min/IP. The
  keyword search keeps its 300/min.

Docs and the OpenAPI document are updated.

Tests

bundle exec rspec is green locally (1531 examples), including the fallback
paths, the cache, and both API surfaces.

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}