Refine coroutine for hart scheduling

[jserv]

This refines I/O coroutine API that was designed for peripheral I/O but provided no benefit over inline polling.

---

Summary by cubic

Refined the coroutine system to focus on hart scheduling and moved peripheral I/O to inline polling for lower latency and simpler design. Added event-driven WFI with a round‑robin scheduler that sleeps when all harts are idle.

• Refactors

  • Removed I/O coroutines; peripherals now inline‑polled every 64 instructions.
  • New API: coro_init(total_slots, hart_slots); added CORO_INVALID_ID. Renamed to coro_create_hart(slot_id, func, arg) and coro_resume_hart(slot_id); coro_is_suspended(slot_id) checks only hart slots.
  • Scheduler drains timer events and monitors UART (kqueue/timerfd + poll) so WFI sleep blocks properly; virtio‑net sets vnet.ram unconditionally and skips queue refresh if peer is not initialized.

• Migration

  • Replace coro_init(n_hart) with coro_init(total_slots, hart_slots).
  • Update calls to coro_create_hart/coro_resume_hart to use slot_id and arg.
  • Use coro_is_suspended(slot_id) only for hart slots; optionally use CORO_INVALID_ID when checking current hart ID.

^{Written for commit cc633b4. Summary will update automatically on new commits.}