perf: skip geometry reconciliation on the text-input redraw

[arzafran]

Closes the remaining half of #183. Measured with the in-process sampler, not the socket harness.

What this does

After every printable keystroke the keyDown path called forceRefresh, which reasserts the surface's display id and runs forceRefreshSurface()'s geometry reconciliation. Both exist for topology changes — split close/reparent, and the stuck-vsync state after wake-from-sleep. A character causes neither: typing changes grid contents, not the surface's size or which display it lives on.

Adds requestRedrawAfterInput(), which only calls ghostty_surface_refresh, and uses it at the text-input call site. forceRefresh is unchanged and still used everywhere else.

Early-out conditions are deliberately identical to forceRefresh's, so when a redraw happens is unchanged — only the work done differs.

The numbers

Two before/after pairs, 400 keystrokes each, timed in-process at the call site:

metric	forceRefresh	requestRedrawAfterInput
p50	0.318 / 0.159 ms	0.004 / 0.003 ms
p95	0.602 / 0.269 ms	0.036 / 0.008 ms
p99	0.726 / 0.369 ms	0.113 / 0.020 ms
mean	0.330 / 0.169 ms	0.011 / 0.004 ms

Complete separation on every metric across both pairs — roughly 40-100x.

Honest about magnitude: the absolute saving is ~0.15-0.3 ms per keystroke. Not transformative on its own, but it is pure waste on the path this codebase treats as sacred, and removing it costs nothing.

Why measuring this took three attempts

The first attempt used tests/test_workspace_churn_up_arrow_lag.py and produced four runs of completely overlapping ranges. That harness measures a socket RPC round trip — IPC, JSON, DispatchQueue.main.sync, AppKit dispatch, then the work. The transport is order-milliseconds; the effect is ~0.3 ms. It was structurally incapable of resolving this, and more samples would not have helped.

Hence #219 (the in-process sampler), #223 (the CI step that runs it) and #224 (its failure diagnostics). #225 fixed a crash those uncovered on the way — the typing profiler terminated the app whenever it was enabled, because eventFields read keyCode on non-key events.

Test plan

• Two before/after pairs, no overlap on any metric
• xcodebuild -scheme programa → ** BUILD SUCCEEDED ** (verified from log contents)
• CI green

Branch is perf/183-m2-fix-a rather than something tidier: the original needed a rebase onto the crash fix, and force-push was correctly denied, so the same single commit went to a fresh name.