perf(api): push recommendations keyset pagination into the store read #981

Follow-up from #974 (keyset pagination for GET /api/recommendations, #723).

Problem

#974 bounds the response per page. Each page request still loads the full filtered recommendation set via ListRecommendations (ListStoredRecommendations has no ORDER BY/LIMIT/cursor), unmarshals it, and sorts it in memory just to return one page. Walking an 8,000-row tenant in 1,000-row pages therefore loads and sorts 8,000 rows eight times, and per-request memory still grows with tenant size.

Suggested fix

Push the keyset boundary (id > cursor) and LIMIT into the store read, with a deterministic ORDER BY. Compute the full-set summary and regions with a separate aggregate query, so page reads stay O(page).

Source: CodeRabbit finding on #974 (thread 4294696025), triaged as out of scope there.