Feat/io wait

[MikiEEE]

Summary

This PR improves SmallOS I/O wait efficiency by keeping readiness registrations alive across scheduler iterations instead of rebuilding the complete polling set before every wait.

• Adds optional Kernel.create_io_wait_set() support while preserving Kernel.io_wait() as the compatibility fallback.
• Uses a persistent selector on Unix, with PollSelector on macOS based on benchmark results.
• Reuses poll()/ipoll() registrations on supported MicroPython platforms.
• Updates registrations only when read/write interests change.
• Removes the duplicate zero-time wait before an idle blocking wait.
• Adds OSList.has_ready() for checking runnable work without dequeuing it.
• Includes deterministic lifecycle tests and an I/O benchmark.
• Adds no runtime dependencies and preserves custom-kernel compatibility.

Performance

Local macOS benchmark results:

• 256 idle descriptors: 91.62 μs → 35.45 μs (2.58x)
• 256 descriptors, one ready: 80.85 μs → 26.46 μs (3.06x)
• Runnable scheduler with one watcher: 20.26 μs → 14.29 μs (1.42x)

Validation

• All 83 unit tests pass.
• Unix and MicroPython-style persistent polling paths are covered.
• Tests cover registration deltas, cancellation, restart, descriptor reuse,
  invalid descriptors, combined masks, error events, and snapshot fallback.
• demos/unix_demo.py passes.
• Package build, wheel smoke test, compilation, documentation indexing, and
  git diff --check pass.

Follow-up

The execution-adapter work lives on a separate branch. Once merged, adapter
completion sources should use the same first/last-source delta registration
path.