process pool: 3 threads per child does not scale past ~hundreds

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

Per-child stdout reader + stderr reader + stdin writer = 3 threads. Fine for tens, painful for hundreds, impractical for thousands.

For the sync v1 we should just document the ceiling. The longer-term fix is probably a tokio-based variant (one runtime, async readers/writers via `tokio::process::Child`) as a separate crate or feature.

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