fix(pg-cloudflare): emit write errors that have no callback

[askalf]

Addresses the diagnosability half of #3757 (it does not claim to fix the hang itself; see below).

Summary

• CloudflareSocket.write() declared callback: (error?: unknown) => void = () => {}. Connection._send() (packages/pg/lib/connection.js:175) calls this.stream.write(buffer) with no callback, so that no-op default swallowed every rejection from WritableStreamDefaultWriter.write().
• A failed write therefore produced no error anywhere: the query sat unsettled until the socket closed on its own, and the application saw only the generic post-close Connection terminated unexpectedly.
• Fix: make callback optional and, when none was supplied, this.emit('error', err) — which is what a net.Socket does and what the rest of pg is already wired for (Connection → Client._handleErrorEvent → _errorAllQueries). Writes that do pass a callback are unchanged and still get the error there.
• Two regression tests next to the existing mock-writer tests in packages/pg-esm-test/pg-cloudflare.test.js: the emit path, and the callback path (so this can't regress the write(data, callback) support from fix: support callback as second argument to CloudflareSocket.write #3747).

$ # --- on base (9683053), regression test present, fix NOT applied ---
$ node --test --conditions=workerd pg-cloudflare.test.js
▶ pg-cloudflare
  ✔ should export CloudflareSocket constructor (8.422557ms)
  ✔ should safely end after the underlying socket has closed (2.069587ms)
  ✔ should call the write(data, callback) callback exactly once (1.215912ms)
  ✖ should emit error when a write without a callback fails (102.766168ms)
  ✔ should report a failed write to the callback rather than emitting error (4.22798ms)
  ✔ should emit close when ending a socket whose closed promise never settles (2.411796ms)
ℹ tests 6
ℹ pass 5
ℹ fail 1
✖ should emit error when a write without a callback fails (102.766168ms)
  Error: error event was not emitted

$ # --- with the fix applied ---
$ node --test --conditions=workerd pg-cloudflare.test.js
▶ pg-cloudflare
  ✔ should export CloudflareSocket constructor (5.316283ms)
  ✔ should safely end after the underlying socket has closed (1.46093ms)
  ✔ should call the write(data, callback) callback exactly once (1.116075ms)
  ✔ should emit error when a write without a callback fails (1.437625ms)
  ✔ should report a failed write to the callback rather than emitting error (0.618606ms)
  ✔ should emit close when ending a socket whose closed promise never settles (0.963671ms)
ℹ tests 6
ℹ pass 6
ℹ fail 0

tsc -p packages/pg-cloudflare/tsconfig.json, prettier --check and eslint on the two files: all clean.

Decisions

• Optional parameter instead of a no-op default. With a default callback the code can't tell "caller wants the error" from "caller passed nothing"; making it optional is what makes the distinction expressible. emit('error') fires only in the branch that was previously a silent drop; every other path is byte-for-byte what it was. end() passes its own callback, so its handling is untouched (the existing close test still passes).
• Not chunking large writes. That was the original hypothesis on g-cloudflare: large single statement (~124KB) reliably kills the connection on Cloudflare Workers - "Connection terminated unexpectedly" #3757, but it can't be proven without a deployed Workers runtime, and @charmander said on the issue they doubt backpressure is the mechanism. Shipping a chunk size would be guessing. This change makes the runtime's actual rejection visible instead, which is the information g-cloudflare: large single statement (~124KB) reliably kills the connection on Cloudflare Workers - "Connection terminated unexpectedly" #3757 is missing: whoever reproduces it on a Worker with this patch should now see the real error rather than "Connection terminated unexpectedly".
• Not touching Connection._send() (pass a callback there): larger blast radius across all transports for a defect in the Workers socket's own contract.
• Not setting writable = false on failure: Client._handleErrorEvent already marks the client unusable once the error is emitted; kept to one change.

Not run here: the full Node × PostgreSQL matrix (needs a live server; this touches only the Workers socket) and the packages/pg/test/cloudflare vitest-pool-workers suite (needs wrangler/workerd). #3718 is the only open PR touching this file; it changes end(), not write(), and is already superseded by #3735.

AI assistance: the bug was traced and the fix and tests were drafted with AI tooling in my workflow, starting from @charmander's comment on #3757; the tests, tsc, prettier and eslint were executed as pasted above. I'm responsible for the change and will handle review feedback.