Fix timerfd handling in epoll reactor

- Consume timerfd on expiry to prevent epoll busy-spinning
  (level-triggered fd must be read to clear readable state)

- Remove last_timerfd_expiry_ caching optimization to match Asio
  (eliminates data race between timer callback and reactor thread)

Author: sgerbino

src/corosio/src/detail/epoll/scheduler.cpp

@@ -588,12 +588,6 @@ update_timerfd() const
 {
     auto nearest = timer_svc_->nearest_expiry();
 
-    // Skip syscall if expiry hasn't changed
-    if (nearest == last_timerfd_expiry_)
-        return;
-
-    last_timerfd_expiry_ = nearest;
-
     itimerspec ts{};
     int flags = 0;
 
@@ -659,6 +653,8 @@ run_reactor(std::unique_lock<std::mutex>& lock)
 
         if (events[i].data.ptr == &timer_fd_)
         {
+            std::uint64_t expirations;
+            [[maybe_unused]] auto r = ::read(timer_fd_, &expirations, sizeof(expirations));
             check_timers = true;
             continue;
         }

src/corosio/src/detail/epoll/scheduler.hpp

@@ -167,8 +167,6 @@ class epoll_scheduler
     // Edge-triggered eventfd state
     mutable std::atomic<bool> eventfd_armed_{false};
 
-    // Track last timerfd expiry to avoid redundant timerfd_settime calls
-    mutable timer_service::time_point last_timerfd_expiry_{timer_service::time_point::max()};
 
     // Sentinel operation for interleaving reactor runs with handler execution.
     // Ensures the reactor runs periodically even when handlers are continuously