pipeline: defer processor close in pool until all workers exit

[twmb]

When pipeline.threads > 1, all worker goroutines share the same processor instances. Previously, when any worker's input channel closed, its loop() defer would call Close() on the shared processors — even while other workers were still mid-processing. This caused errors like "sql: database is closed".

Add a noCloseProcs flag to Processor so that pool workers skip closing processors in their defer. Instead, Pool.loop() closes the shared processors after all workers have exited.

Fixes redpanda-data/connect#1655

[claude]

Commits
LGTM

Review
The fix correctly prevents early closure of shared processors when a pool worker exits before others finish. The Pool now owns processor lifecycle, waiting for all workers via WaitForClose before closing processors. Test coverage validates the race condition scenario.

LGTM