fix(net): report websocket send failures instead of panicking

WebsocketConnection::enqueue_write_op unwrapped the result of
start_send, which fails whenever the sink has left the active state.
Since the readiness probe in may_enqueue_more_ops doesn't check the
sink state (tokio-websockets returns Ready without inspecting it
below the flush threshold), a server initiating the closing handshake
raced with any enqueued op — including the periodic PING — and
panicked the connection handler, silently killing the client without
a reconnect.

Store the failure and report it once through poll_read_next /
poll_flush so the handler observes it as a regular connection error
and recycles the connection. New ops are rejected while the error is
pending.