fix(engine.io): run the polling write callback when the client aborts a compressed response

[chuanghiduoc]

Problem

Since 91dd763 ("perf(eio): stream compressed polling responses"), doWrite() invokes its callback only from the response's 'finish' handler. If the client aborts the polling request while a large compressed payload is still streaming, 'finish' never fires — and nothing errors on the zlib pipeline either, because writes into the destroyed socket just become no-ops.

Consequences of the lost callback (write() in polling.ts):

• req.cleanup() never runs for that request
• this.emit("drain") never fires, so the send callbacks queued for that batch (sentCallbackFn) are dropped

Before 91dd763, the whole payload was buffered and the callback ran unconditionally right after res.end(), so this could not happen.

Reproduction: new test packages/engine.io/test/compression-abort.js — a client opens a polling request with Accept-Encoding: gzip, the server keeps sending incompressible random data (threshold 0), and the client destroys the request as soon as the response headers arrive. On current main, the server-side send callback of that batch is never invoked; with this PR it always is (verified by stashing the fix and re-running).

Fix

The callback now also runs on the response's 'close' event, with a guard so it still runs exactly once ('close' also fires after normal completion):

this.res.once("finish", done);
this.res.once("close", done);

Testing

• new test: packages/engine.io/test/compression-abort.js (fails without the fix, passes with it)
• existing suites unaffected