[BUGFIX] Ring: fix DoBatch cleanup callback blocked forever on callback panic

[sandy2008]

What this PR does

ring.DoBatch submitted closures called wg.Done() inline after callback() and tracker.record(). If callback panicked, wg.Done() was skipped, causing wg.Wait() to block forever and the cleanup callback to never execute.

Symptom: any panic inside a DoBatch callback permanently blocked the cleanup goroutine. For the distributor, this leaked the context.WithTimeout timer (until RemoteTimeout expired) and request buffers (req.Timeseries, req.Free()) — they were never reclaimed. The same issue affected all three production DoBatch callers: distributor, alertmanager distributor, and multitenant alertmanager.

Fix: move wg.Done() to defer wg.Done() so it runs even during panic unwinding, ensuring the cleanup goroutine always completes. An earlier iteration also added defer cancel() to the distributor's doBatch function, but this was intentionally reverted after review found it contradicts the design intent: "Use a background context to make sure all ingesters get samples even if we return early." The defer wg.Done() fix makes the existing cleanup callback (which already calls cancel()) reliable on all paths, preserving send semantics without cancelling in-flight ingester RPCs.

Which issue(s) this PR fixes

Fixes #7558

Checklist

• CHANGELOG.md updated — not yet, will add if maintainers confirm the fix direction.
• Documentation updated — not applicable, no flags or config changed.
• Tests added — TestDoBatchCleanupCalledOnCallbackPanic verifies cleanup runs even when callback panics.

Test plan

• go vet ./pkg/ring/... ./pkg/distributor/... — clean
• go test -tags "netgo slicelabels" -count=1 -timeout 120s ./pkg/ring/... — 10/10 packages pass
• go test -tags "netgo slicelabels" -run TestDistributor_BatchTimeoutMetric ./pkg/distributor/... — passes
• New test passes with fix, fails without (verified by reverting defer during development)

[friedrichg]

Thanks. the test proves to me this is needed