Hand rolling

[Ataba29]

What

Replaces the thread-per-connection model with a cross-platform,
readiness-based event loop (epoll on Linux, IOCP on Windows), unified
behind an IEventLoop interface.

Why

Previously, every connected client permanently occupied one thread-pool
worker for its entire lifetime, capping concurrent connections at the
pool size (3) regardless of how idle those clients were. Now a small
number of threads watch all sockets for readiness; only actual command
execution (GET/INSERT/DELETE) uses the thread pool.

Changes

• IEventLoop interface with EpollEventLoop (Linux) and IocpEventLoop
  (Windows) implementations, selected at compile time via
  EventLoopFactory::makeEventLoop()
• Connection struct + Server::connections map, keyed by socket
• acceptClients() now sets sockets non-blocking and registers them
  with the event loop instead of spawning a thread per client
• messageHandler() now does one recv() per event instead of looping,
  and correctly treats EWOULDBLOCK as "nothing to do" rather than a
  disconnect
• Added a busySockets guard to stop duplicate recv() jobs from being
  queued for the same socket before a prior job finishes (level-triggered
  epoll can otherwise report the same socket ready multiple times before
  it's drained)
• wait() now times out after 1s instead of blocking forever, so
  graceful shutdown (stop()) doesn't hang waiting on client activity
• Fixed test_ratelimiter.cpp, as in previous updates a change to the isAllowed() happened.

Testing

Manually tested with multiple concurrent ncat clients: INSERT/GET/DELETE
all confirmed working with 2+ simultaneous connections, no spurious
disconnects. Graceful stop verified to return promptly with clients
still connected, as well as running unit tests and having all of the 16 tests pass.

Known follow-up (not in this PR)

UserSession::terminate_session() closes the client socket directly when
a session idle-times-out, without telling Server — so connections and
the event loop can end up referencing a socket that's already been closed
elsewhere. Needs to be routed through Server::closeConnection() for a
single consistent teardown path. Filing as a separate follow-up rather
than expanding scope here.

[Ataba29]

Ran stress test and here are the results

Metric	200 clients	2,000 clients	10,000 clients
Connected	200/200	2,000/2,000	10,000/10,000
Requests completed	4,000/4,000	40,000/40,000	199,980/200,000
Throughput	11,908 req/sec	10,401 req/sec	8,207 req/sec
p95 latency	15.5 ms	190.7 ms	1,199 ms
Errors	0	0	1 timeout