fix(rsqueue): buffer PollAddress errCh to prevent goroutine leak on abandoned poll

[jonyoder]

Summary

DatabaseQueue.PollAddress returned an unbuffered error channel. Its polling goroutine has exactly one send (errCh <- err on a non-lock store error). When the caller stops receiving — e.g. its request context is canceled and it returns from the select { case <-ctx.Done(): ... case err = <-errCh: } wait in rscache/file.go — that send parks forever, leaking one goroutine per abandoned poll. Under bulk request pile-ups these accumulate and degrade the server.

Downstream this is rstudio/package-manager#19008 (blocked PollAddress senders observed in production goroutine dumps).

Fix

Buffer the channel by 1 (make(chan error, 1)). The goroutine sends at most one value before closing, so a single slot guarantees the send never blocks regardless of whether a receiver is still present. This mirrors the "guarantee the sender can always complete" approach rather than depending on a live receiver.

• No behavior change for a live waiter: it still reads the single error (or the close) exactly as before. Errors are neither newly surfaced nor swallowed.
• The only production caller is rscache/file.go (fileCache.get/Get); RecordingProducer.PollAddress is a test double and unaffected.

Test

Adds TestPollErrAbandonedReceiver — starts a poll, never receives from the returned channel, and asserts (via leaktest) the poll goroutine has exited. RED before the fix (leaked goroutine parked at the send), GREEN after.

Note: the new test uses the correct defer leaktest.Check(c)() form. The existing poll tests call defer leaktest.Check(c) without invoking the returned checker, so their leak assertion is currently a no-op — flagged here; fixing those is out of scope for this change.

go test ./pkg/rsqueue/... ./pkg/rscache/... and go vet pass.

Downstream

Package Manager (#19008) vendors this module; once released (a patch tag, e.g. v4.3.1), PPM will bump go.mod + go mod vendor and add a ### Fixed NEWS entry referencing #19008.