fix(socketio): clear the 5s emit timer on acknowledgement

[RobertoReale]

Fixes #405

Problem

The promisified emit starts a 5 s timer per call and never clears it:

const timeoutPromise = new Promise((_, reject) => {
    setTimeout(() => { reject('timeout'); }, 5000);   // never cleared
});

When the acknowledgement arrives the race is already settled, so the late reject is a no-op and nothing is visible — but a timer stays pending for 5 s after every emit and still runs its callback. On a path that runs continuously while editing (joinDoc, applyOtUpdate, clientTracking.*) that is one live timer per call, and it makes the timeout impossible to instrument: adding any log to that callback reports a failure for every successful call, 5 s late. That is exactly what happened while I was debugging this extension — the console filled with 'applyOtUpdate' timed out after 5s for saves that had been acknowledged in ~150 ms and had already appeared on overleaf.com.

Two related points, both in the same few lines:

• init() replaces this.socket when the handshake scheme changes (v1 → v2 on overleaf.com). A call emitted on the previous socket can never be acknowledged, so its timer always fires even though the call was abandoned by design; binding the call to the socket it was emitted on tells the two cases apart.
• when the server rejects an emit, the raw payload is passed on and VS Code renders it as Unable to write file ... ([object Object]), hiding the actual message.

Change

• clearTimeout in the acknowledgement callback;
• capture socketAtEmit and emit on it, so a timeout on a socket that has since been replaced is not reported as a failure;
• log a rejected emit through a small stringifyError helper (name/code/message, falling back to JSON), instead of letting [object Object] reach the user.

Verification

Against overleaf.com with 0.15.10 + this patch, tracing socket.io 0.9 packets: every emit that asks for an ack (5:<id>+::…) gets its ack ackId=<id> back, and no timeout is logged any more — before the fix the same session produced ~20 spurious timeout lines. The one call that legitimately never completes (the joinProject queued on the v1 socket that overleaf.com rejects) is now silent instead of printing an error.