process pool: I/O thread errors vanish silently

Came up while planning the optative-process-pool extraction.

The stdin writer thread breaks on `write_all` failure without telling anyone. The user keeps calling the per-child sender and either gets swallowed-success (if the channel still accepts) or a generic SendError that doesn't say why the child is no longer reachable.

At minimum log via `tracing` when an I/O thread exits abnormally, with the process key and the error. Ideally route the error to a side-channel the consumer can observe.

Drafted by Claude (AI assistant) during a design review.