Improve sleep performance and precision

[mscheltienne]

I got Claude to work, I got no clue what it did, I'll let you dissect the changes, have a look:

  Results                                                                                                                               
                                                                                                                                        
  ┌──────────┬──────────┬───────────┬───────────┬──────────────────────────────────────────┐                                            
  │ Duration │  Before  │   After   │  stimuli  │               Improvement                │                                            
  ├──────────┼──────────┼───────────┼───────────┼──────────────────────────────────────────┤
  │ 1μs      │ 4.07μs   │ 1.05μs    │ 1.65μs    │ 3.9x faster (now beats stimuli)          │
  ├──────────┼──────────┼───────────┼───────────┼──────────────────────────────────────────┤
  │ 10μs     │ 12.20μs  │ 10.08μs   │ 10.77μs   │ 1.2x faster (now beats stimuli)          │
  ├──────────┼──────────┼───────────┼───────────┼──────────────────────────────────────────┤
  │ 100μs    │ 102.08μs │ 100.11μs  │ 100.91μs  │ tighter                                  │
  ├──────────┼──────────┼───────────┼───────────┼──────────────────────────────────────────┤
  │ 1ms      │ 1154μs   │ 1000.64μs │ 1003.40μs │ 153μs less overshoot (now beats stimuli) │
  └──────────┴──────────┴───────────┴───────────┴──────────────────────────────────────────┘

  What changed in src/lib.rs

  Three conceptual fixes to the sleepers function:

  1. Instant::now() instead of SystemTime::now() — monotonic clock, appropriate for elapsed time measurement, uses mach_absolute_time()
  on macOS with ~42ns resolution on Apple Silicon
  2. Adaptive halving instead of single coarse sleep — sleep for half the remaining time when >200μs remains, converging safely without
  overshooting
  3. std::hint::spin_loop() instead of thread::sleep(500ns) — pure CPU spin-wait hint (ARM YIELD instruction) for the final <200μs,
  avoiding the OS scheduler entirely