feat(api): add ParadeDB pg_search as Citus-compatible BM25 backend

[nicoloboschi]

Summary

Adds a fourth value (pg_search) for HINDSIGHT_API_TEXT_SEARCH_EXTENSION alongside the existing native, vchord, and pg_textsearch. ParadeDB pg_search is the only true-BM25 backend that works on a Citus distributed Postgres cluster — both vchord_bm25 and Timescale pg_textsearch are Citus-incompatible — so this unblocks horizontally scaled deployments. Closes #1754.

The change slots into the existing pluggable design: every backend-specific branch lives behind if config.text_search_extension == ..., so this PR just adds a fourth arm wherever the existing three live. No new abstraction is introduced.

Notes on the SQL shape

• The retrieval arm builds the @@@ predicate as paradedb.boolean(should => ARRAY[paradedb.match('text', \$4), paradedb.match('context', \$4), paradedb.match('text_signals', \$4)]) rather than the simpler id @@@ \$4. The @@@ operator on the key_field requires field-qualified terms (text:foo), and the boolean fan-out keeps multi-field coverage without query-string interpolation.
• The runtime reconciler (ensure_text_search_extension) disambiguates pg_textsearch vs pg_search by inspecting pg_indexes.indexdef for the key_field reloption — both extensions register a bm25 access method on a text column, so the access-method check alone is ambiguous.
• INSERT path is a no-op for pg_search (same as pg_textsearch); the BM25 index reads base columns directly.

What ships

• Config validation accepts pg_search
• engine/sql/postgresql.py build_bm25_arm / prepare_bm25_text branches
• Alembic migration arms for memory_units, learnings, reflections, and the text_signals follow-up migration
• Runtime reconciler arm in migrations.py (incl. pg_textsearch vs pg_search disambiguation)
• Docker compose example based on paradedb/paradedb:latest-pg17
• Docs in configuration.md and retrieval.md
• Unit tests for the new SQL shape

Test plan

• uv run pytest tests/test_db_abstraction.py tests/test_migration_shape.py tests/test_admin_backup_restore.py — 151 passed
• ./scripts/hooks/lint.sh — passes
• uv run ty check hindsight_api/ — passes
• End-to-end validation against the docker-compose example:
  • Built the ParadeDB image and ran migrations to head
  • Verified the BM25 index DDL is exactly USING bm25 (id, text, context, text_signals) WITH (key_field=id)
  • Retained four memories via POST /memories
  • Recall with query "BM25 ranking ParadeDB" returned the ParadeDB row at rank Fix dependencies and docker #1, bank isolation respected
• Reviewer to spot-check the boolean-fan-out query shape — happy to switch to a simpler text @@@ \$4 form scoped to a single column if that's preferred
• Optional follow-up (not in this PR): bring up a Citus coordinator + worker with pg_search, distribute memory_units on bank_id, confirm recall still works