feat(perf): evaluation caching (Caffeine) + ETag/304 on SDK endpoint

[trinhvandat]

Context (P1 — scale)

Every SDK GET /api/v1/sdk/flags call hits PostgreSQL. Polling SDK fleets multiply this. Redis is on the v2 roadmap; Caffeine covers single-instance now.

Scope

• Caffeine cache keyed by environment id over the evaluation response; invalidate on flag/flag-state/flag-archive mutations (and on flag creation which fans out states)
• ETag (hash of payload or env-version counter) + If-None-Match → 304 handling on the SDK endpoint
• Cache metrics via Micrometer (hit ratio)
• Config: TTL + max size in properties

Acceptance criteria

• Repeated evaluation calls do not hit the DB (verify via query counter or datasource-proxy in test)
• Any admin mutation to a flag/state invalidates that environment's entry (integration test)
• Clients sending If-None-Match get 304 with empty body