fix(output): drain buffered records on TCP/UDP Stop (PIPE-1230) - #271

Proposed Change

The TCP and UDP outputs could drop buffered records on Stop(). Stop() closed the data channel and cancelled the worker context at the same time. The worker's select then chose at random between draining the channel and taking the cancelled context, so it often exited with records still buffered.

Stop() now stops the workers first, then drains whatever is still buffered before returning. With the workers gone the drain is the only reader, so delivery is deterministic. The drain is bounded by the write timeout, the stop timeout, and the caller's context. An unreachable destination can't hang shutdown.

How to validate

• Automated: go test -race ./output/tcp/ ./output/udp/ passes. New tests cover the drain (records buffered at Stop all arrive), an unreachable destination (Stop returns without hanging), and the send-failure and context-cancelled paths. drainBuffered and drainTo are at 100% coverage.
• Manual: point a TCP output at a listener, write a burst of records, call Stop, and confirm the listener receives all of them.

Checklist

• Changes are tested
• CI has passed