Fix sentiment 502 and cut per-request latency across slow endpoints

[nabindev3]

The /api/v1/sentiment/analyse endpoint was returning 502 Bad Gateway from Render's proxy because the engine could spend up to ~400s per request in the worst case. Several other endpoints had silent slowness from re-reading CSVs and rewriting cache files on every request.

Sentiment engine (src/hf_sentiment_engine.py, src/sentiment_engine.py):

• Replace per-request disk cache (full JSON read + full JSON rewrite on every call) with a thread-safe in-memory LRU. Disk cache was O(N) I/O per request and burned Render's ephemeral disk.
• Cap HFInferenceClient TIMEOUT 25s->8s, RETRY_ATTEMPTS 3->1, COLD_START_WAIT 20s->3s, and disable wait_for_model so HF returns 503 fast on a cold model instead of blocking up to 20s.
• Fan out the 3 HF model calls (sentiment, irony, zero-shot aspect) concurrently with a ThreadPoolExecutor. Latency is now bounded by the slowest of the three, not their sum.
• Drop the 0.3-0.5s sleep between batch items; replace sequential batch with bounded concurrent execution.
• Don't cache failure results so the next request can retry the warm engine.
• Include engine identifier in cache key so a TextBlob fallback doesn't poison later warm-engine lookups.

Sentiment router (backend/routers/sentiment.py):

• Hard endpoint deadline (12s single, scaled+capped 45s batch). If the underlying engine doesn't return in time we serve a TextBlob result instead of letting the request hang until Render's proxy emits 502.

Hot-path I/O caching:

• backend/routers/forecast.py: add @lru_cache to _load_daily and _load_ext (were re-parsing 2192-row + 6-year CSVs on every /forecast call).
• backend/routers/pricing.py: same fix for daily_kpis and external_regs.
• backend/routers/xai.py: cache the 180k-row bookings DataFrame used by /global-importance instead of re-reading it per request.

All 25 existing tests pass; suite runtime dropped from 5.0s to 3.9s.