Fix Nonblocking writes

[Zeleph25]

Add non-blocking write buffer to NetQueue

Problem: write_n_bytes spun with usleep(200) on EAGAIN. Since this runs inside dispatch handlers inside wait_and_process, send buffer pressure on any single connection stalled the entire event loop. Dropping on EAGAIN wasn't an option — losing CONNECT_TO_SWITCH or AGENT_CONNECTION_TOKEN breaks the handshake.

Fix: Each NetworkElement gets an outbound write buffer. On EAGAIN mid-write, remaining bytes are saved into the buffer and EPOLLOUT is registered on the fd. When the kernel send buffer drains, epoll fires and flush_write_buffer drains the queue. Once empty, EPOLLOUT is deregistered. Connections exceeding 64KB buffered are marked SCK_ERROR and closed on the next loop iteration — a connection that full isn't reading.

All in-loop sends in conductor, switch, and agent go through queue->send and queue->send_heartbeat. Blocking MIP phase sends stay on send_protobuf_packet since those fds aren't in the queue yet and are still in blocking mode.

[TheoHollender]

May be interesting to do some form of garbage collection at some point if the buffer is too big, but for now should be fine.