feat(vt): add SafeEmulator.View/Update for batched locked access

[LXXero]

Problem

SafeEmulator.CellAt takes an RLock (plus its deferred unlock) on every call. Any consumer that walks the cell grid per frame — i.e., every renderer built on vt — pays this per cell: an 80×24 grid is ~2k lock acquisitions per pass, and real applications run multiple passes per frame on much larger grids.

Profiling xerotty (a terminal emulator built on vt) under continuous redraw with Linux perf: after eliminating our own per-cell costs, SafeEmulator.CellAt + its defer wrapper remained the top CPU entry at ~16% of total process time (~15k lock+defer pairs per frame on a 240×65 grid at 20fps ≈ 300k lock ops/sec, all to read cells that cannot change mid-frame).

Change

bbolt-style scoped accessors, purely additive:

func (se *SafeEmulator) View(fn func(*Emulator))   // batched reads under one RLock
func (se *SafeEmulator) Update(fn func(*Emulator)) // batched writes under one Lock

Renderers wrap each frame's grid walk in one View; bulk mutations (e.g. replaying a snapshot via SetCell) batch under one Update. Existing per-call methods unchanged.

Benchmark

An 80×24 grid walk, per-cell CellAt vs the same reads under one View:

BenchmarkCellAtGridWalk-16     162933    14081 ns/op
BenchmarkViewGridWalk-16      1885207     1255 ns/op

11× — and the gap widens with grid size (lock cost is per-cell, savings are per-batch).

Testing

• TestSafeEmulatorViewUpdate — basic read/write through the accessors
• TestSafeEmulatorViewConcurrent — grid-walking Views against concurrent Write pressure, -race clean
• Benchmarks above included

Related: #879 (same SafeEmulator synchronization layer).