fix(tui): coalesce part stream deltas per frame to prevent stalls - #41472

Problem

Clicking to expand a streaming <thinking> block in the TUI can stall the UI until reasoning finishes.

Root cause chain:

1. The LLM stream emits session.next.reasoning.delta (and text.delta / tool.input.delta) events per chunk, unbuffered (packages/core/src/session/runner/publish-llm-event.ts:278).
2. The TUI store applied every delta synchronously: match.text += event.data.delta on each event (packages/tui/src/context/data.tsx).
3. While a reasoning block is expanded, every store write re-renders its markdown block and OpenTUI re-parses + re-highlights the entire accumulated text on each token — O(N) work per delta, O(N²) over the stream. Long thinking blocks (thousands of tokens) pin the render loop and the TUI appears frozen.

Note the web app already avoids this exact problem with per-part delta accumulation (part_text_accum_delta, #26822); the TUI had no equivalent.

Change

Adds a per-frame delta coalescer (packages/tui/src/context/delta-buffer.ts):

• Deltas accumulate per (kind, sessionID, messageID, partID); the store is rewritten at most once per animation frame instead of once per event.
• drain() runs before any non-delta event (e.g. *.ended) so terminating events that replace the full part text still observe the final accumulated deltas; also drains on provider unmount.
• Covers text.delta, reasoning.delta, and tool.input.delta (identical pattern, same fix).

Testing

• New unit tests for createDeltaBuffer (coalescing, per-part isolation, drain semantics, stale flush no-op, single-schedule). bun test passes 5/5.
• Could not run tsgo --noEmit in this environment (tool not installed); changes are type-annotated to match the module's existing usage.
• Not yet manually verified in a live TUI session.

Follow-up (not included)

• The v2 session durable-store path (packages/core/src/session/message-updater.ts:358) performs the same per-delta full-text rewrite; worth batching there separately.