Add getFrameController() for application-owned text selection

[chiga0]

This is the first part of the split suggested in #980: a small, opt-in, read-only frame controller for apps that own their alternate-screen viewport and implement selection themselves. Selection semantics for Text (flows, boundaries, selectable and friends) are left for a follow-up once the bridge is proven by a real consumer.

What's included

• getFrameController(stdout) with:
  • getFrame() — a read-only cells[y][x] grid; each cell exposes only value + fullWidth. Wide characters occupy two cells: the leading cell has fullWidth: true, the trailing placeholder has an empty value. No internal style data is exposed.
  • getSelection() / setSelection() — Ink highlights the region before serialization. Repaints go through the regular render throttle, and identical selections are deduplicated.
  • subscribe() — frame listener, returning an unsubscribe function.
• Selection coordinates are zero-based cell positions in Ink's output region, normalized to reading order — reverse (right-to-left or bottom-to-top) drags select the same region as forward drags.
• Readme docs covering the API and coordinate semantics.

Concerns from #980 addressed

1. Per-render overhead — cells are only projected while subscribers exist, so apps that never use the frame controller pay no overhead.
2. Subscriber reentrancy — listeners run in a microtask after the publishing render completes; a setSelection() call from a listener schedules a new render instead of re-entering the one in flight; notifications are coalesced and frames delivered in order.
3. API surface — frame publishing stays internal (getFrameController returns the read-only view); no internal style data is exposed; frames and rows are frozen and deeply readonly-typed.
4. Reverse selections — normalized in setSelection(), covered by tests.
5. Tests — 14 tests covering opt-in behavior, async delivery, wide characters, multi-row regions, reverse selections, dedupe, and subscriber reentrancy. The nested-text metadata concern from Add text selection system with frame-level cell composition #980 doesn't apply here because this PR adds no semantic Text props.

Also included

renderInteractiveFrame re-asserts the app's current cursor intent before each interactive render: log-update only applies a cursor position set since the previous render, so without this a repaint triggered by setSelection() would move the cursor from its useCursor() position to the end of the output.

The alternate-screen clearTerminal from #980 is deliberately not included here: it's an unrelated behavior change and can be a separate proposal if desired.