fix(mobile): cut chat re-renders, list churn, and store I/O

[snehasishcodes]

Batch #2 perf fixes from mobile deep audit.

Stores:

• chat persist excludes draftBySession (was rewriting AsyncStorage per keystroke)
• messages capped at 500 per session (was unbounded with base64 URLs)
• models fetch deduped with inflight guard + 5min TTL + deeper name/provider compare

Chat screen:

• Stable zustand selector factories via useMemo (was useCallback creating new fn per render and resubscribing)
• connectionStatus isolated into ConnectionBanner (was re-rendering full 1000-line screen on SSE status flips)
• modelByAgent prop memoized (was new object per render breaking ChatInput memo)
• Pagination offset excludes local-/error- optimistic ids; initial merge sorted by created
• Question poll backed off 5s to 30s fallback (SSE already pushes same events)
• FlatList tuned with updateCellsBatchingPeriod/disableIntervalMomentum and onEndReached disabled while streaming

Lists:

• sessions now prop removed; rows use 5min-bucketed useNowBucket instead of all-rows re-render per minute
• MessageMetadata uses getState lookup (was subscribing every row to full arrays); MemoPartRenderer compares id/parts.length instead of object ref (was re-rendering all parts per token)
• CodeBlock memoized, prism grammars deferred off render, 20k char truncate with show-more, copy timer cleanup.