Fix broken cursor pagination for word_similarity relevance search

[Petah1]

Summary

closes #194. Cursor pagination for searchCampaignsByRelevance() / searchDonationsByRelevance() in src/services/search.service.ts had two real bugs, not just the float-precision issue described in the issue:

• A hard crash: the WHERE clause referenced the SELECT-list alias score directly (WHERE (score, id) < (...)). Postgres does not allow referencing an output alias from the same query's WHERE clause — verified against a live Postgres 15 instance that any second-page request threw column "score" does not exist.
• A type-promotion bug: word_similarity() returns float4. The cursor's score, once round-tripped through Prisma into a JS number and sent back as a query parameter, gets bound as numeric. Comparing a float4 column against a numeric parameter silently produces wrong results — confirmed directly against Postgres that an equality check between a real column and its own round-tripped value returned zero matching rows.

Fix

• Cast the score expression to double precision everywhere it's used (SELECT, > 0.2 threshold filter, cursor comparison), so every comparison happens in one consistent Postgres type instead of relying on implicit float4/numeric promotion.
• Replace the tuple comparison (score, id) < (x, y) with an explicit score < x OR (score = x AND id < y), which also removes the alias-in-WHERE dependency causing the crash.
• Add COALESCE(word_similarity(query, description), 0) for campaigns, matching the existing NULL-handling already used in donation search.
• Version the cursor payload (v: 2). A pre-fix (unversioned) cursor is now detected and treated as absent — falls back to page one — instead of being silently misinterpreted with the old, broken comparison semantics.

Applied identically to both searchCampaignsByRelevance() and searchDonationsByRelevance(). Beneficiary search and the non-relevance sort paths are untouched (out of scope per the issue).

Test plan

• Unit tests (mocked Prisma) in src/services/search.service.test.ts: cursor encode/decode roundtrip for a float4-range value, legacy/unversioned cursor is dropped gracefully, adversarial manipulated cursor (score: -1) doesn't crash or leak rows, generated SQL uses CAST(... AS DOUBLE PRECISION) and COALESCE for description.
• Integration tests in tests/performance/search.performance.test.ts against a real Postgres 15 + pg_trgm instance (validated locally with docker):
  • 50-campaign match set, 10/page: all 50 unique IDs returned, no duplicates, no gaps.
  • 20 campaigns with an identical word_similarity score: partition cleanly across two pages by id DESC, zero overlap.
  • Manipulated cursor score returns empty results without crashing.
• npx tsc --noEmit clean for the changed file.
• Existing cursor-based-pagination unit tests still pass (updated to build v2-format cursors). Pre-existing unrelated test failures in the same file (offset-pagination expectations for a different code path, tiebreaker-order mismatches, unrelated mock gaps in searchBeneficiaries/globalSearch tests) predate this change and are unaffected.