feat: same-turn memory retrieval in provider mode

prefetch() now reuses the in-flight background search when it matches
the current query, or falls back to a synchronous search (capped at 3s,
split across both requests in case-insensitive mode) so memories are
injected on the same turn instead of one turn behind.

- Track _prefetch_query to scope cached results to their query
- Reset _prefetch_result when starting a new background search to
  prevent stale cross-turn injection after a timed-out join
- Guard background thread writes against superseded queries
- Add optional per-request timeout override to LocalMem0Client
- Update AGENTS.md and README to describe the same-turn flow

AGENTS.md

@@ -43,8 +43,8 @@ curl http://localhost:8000/health
 
 **Memory injection flow:**
 1. `queue_prefetch()` — Spawns background thread on user message
-2. `prefetch()` — Returns cached result next turn
-3. `pre_llm_call` hook — Sync prefetch for immediate injection (fallback context)
+2. `prefetch()` — Sync retrieval for the current query (same-turn injection); falls back to cached background result when no query is provided
+3. `pre_llm_call` hook — Sync prefetch for immediate injection (plugin mode)
 4. `sync_turn()` — Non-blocking server-side fact extraction
 
 **Circuit breaker:** 5 consecutive failures → 120s cooldown.