Add enforced scheduler safepoints and timer fairness

[drpaneas]

Summary

The current scheduler is cooperative by convention. Long-running Go code must call runtime.Gosched() manually, and timer progress depends on returning to the scheduler loop.

Why

• A CPU-bound goroutine can starve other goroutines and delay timer wakeups indefinitely.
• Fairness and frame pacing become application-discipline problems instead of runtime guarantees.
• This is a practical usability issue even if the runtime stays M:1.

Evidence

• runtime/runtime_stubs.c: runtime.Gosched() is documented as something callers must do in long-running loops.
• runtime/scheduler.c: goroutines run until they yield, block, or exit.
• runtime/scheduler.c: check_timers() is called from scheduler_run_loop() only after control returns to the scheduler.

Direction

• Keep the single-threaded scheduler, but add enforced safepoints or deferred preemption at safe runtime boundaries.
• Make timer servicing part of a fairness model, not just the outer scheduler loop.
• Add a budgeted scheduling mode that can cooperate with frame timing and vblank-driven runtimes.