perf: bound terminal output scheduling

[dliu120]

Why

Large PTY reads can monopolize the main delivery queue, while interactive output and hidden terminal views can continue presentation work in the same main-loop turn. Both reduce the time available for input handling during heavy terminal output.

What changed

• Deliver queued PTY output in ordered slices of at most 16 KiB while preserving the existing read size and backpressure thresholds.
• Coalesce interactive display requests through the existing pending-display gate, including feeds already running on the main thread.
• Add an explicit presentation-active state: parsing continues while hidden, but display, accessibility, Metal, cursor/text blink, and related callbacks pause.
• Generation-gate queued presentation callbacks so stale work cannot resume after visibility transitions.
• Mark the full screen dirty and schedule one refresh when presentation resumes.
• Preserve synchronized-output behavior.
• Remove a duplicate debug-display update from the display wrapper.
• Add focused coverage for byte ordering, delivery bounds, queue compaction, interactive display coalescing, and presentation suspension/resume.

Testing

• swift test --filter LocalProcessSchedulingTests
• focused presentation/scheduling coverage: 7/7
• swift test (701 tests across 60 suites)
• git diff --check
• CI green

Notes

This reduces and suspends bounded presentation work but does not claim a measured end-to-end latency improvement without runtime traces.

[migueldeicaza]

I redid the IO system in SwiftTerm in this branch, which will soon become the main branch (v2), please take it for a spin as it touches the same issues:

new-io-perf-glyph-cache