Reduced CPU usage

[g2px1]

Summary

Performance and correctness fixes across the timer wheel, accept path, and coroutine dispatch.

---

Changes

perf(timer): replace std::list buckets with std::vector, fix O(n) scan in updateNextExpiryTime

Timer wheel buckets used std::list<Timer*>, causing pointer chasing on every tick iteration.
Replaced with std::vector<Timer*> and swap-and-pop removal — O(1), order within a bucket is irrelevant.

updateNextExpiryTime was called after every advance() — up to N times per tick, each doing a full
scan of all 1024 buckets. Fixed by maintaining nextExpiryTime_ incrementally in addTimerToWheel
and invalidating it to 0 in removeTimerFromWheel. Full rescan now runs only when the minimum is
unknown (after removal or expiry of the minimum timer). Added early return for elapsed == 0.

Flamegraph before: ~85% CPU in TimerWheel::tick
Flamegraph after:  ~1.4%, on par with epoll and coroutine dispatch

---

feat(net): add async_accept(F, Args...) overload, deprecate async_accept()

The zero-argument async_accept() accepted one connection per ET wakeup and returned. In
edge-triggered mode (epoll ET / kqueue EV_CLEAR) this silently drops all connections buffered
since the last edge.

New overload drains the accept queue in a tight loop until EAGAIN before suspending — one suspend
per batch of connections instead of one per socket. F may be a plain void callback or a
coroutine function; coroutines are automatically spawned via co_spawn. Extra arguments are
forwarded to F on each call via variadic Args....

Implemented independently per backend:

Backend	Strategy
Linux epoll	Drain loop + AwaiterAccept on EAGAIN
Linux io_uring	Try accept4 first, fall back to AcceptAwaiter SQE
BSD / macOS kqueue	Same drain strategy, manual fcntl(O_NONBLOCK|FD_CLOEXEC)
Windows IOCP	Full AcceptEx loop inlined, no recursive async_accept() call

Old overload preserved with [[deprecated]] for backwards compatibility.

---

fix(system): correct promise cast in co_spawn_static overloads

Template overload cast promise to AwaitableFrameBase by value, slicing the object —
set_thread_id was never called on the original. Fixed with static_cast<AwaitableFrameBase*>.

Handle overload had a malformed cast on coroutine_handle<> — erased handles cannot be cast
directly. Fixed by restoring the typed handle via from_address before accessing the promise.

---

Testing

Profiled under wrk load (-t8 -c500 -d30s) with perf record -F 999. Flamegraph attached.
No regressions in existing sync primitives, channels, or timer tests.