Add cross-page notification delivery reliability analytics

### Problem

The alert delivery page supports per-record inspection and retry, but its channel health is hidden by pagination. Operators cannot tell whether a failure is isolated or channel-wide, quantify delivery latency, find records stuck in an in-flight state, or group recurring provider errors without manually exporting many pages.

### Proposed change

Add a read-only reliability analysis that loads delivery history through the existing paginated endpoint (capped at the latest 10,000 records), then provides:

- terminal success rate, P95 delivery latency, retry and stalled-delivery counts;
- per-channel success, failure, in-flight, retry, average latency, and P95 latency;
- normalized recurring error signatures with affected alert counts;
- searchable/filterable delivery details and CSV export;
- explicit truncation disclosure plus Chinese/English copy and focused tests.

The analysis must not trigger retries or modify delivery records.

Start at the alert delivery page and trace the existing paginated delivery-history endpoint. Define the read-only analysis around the 10,000-record cap, then review the focused tests needed for metrics, channel breakdowns, error signatures, filtering, CSV export, truncation disclosure, and Chinese/English copy.

Written by the indexing model from the issue text.