scheduler: make closed-stream cleanup expiry-driven

[anmonteiro]

Summary

• decouple closed-stream cleanup from write-path polling
• replace per-poll marked_for_removal table scans with an expiry queue keyed by stream id
• keep cleanup work proportional to actually-expired entries on each scheduler tick
• tick cleanup in both client/server read polling paths
• add regression test showing closed streams are reaped even without next_write_operation polling

Details

Previously, TTL decrement/removal for closed streams happened inside Scheduler.flush (write scheduling), so idle/read-mostly connections could retain closed streams indefinitely if writes were not polled.

This PR introduces:

• cleanup_epoch + cleanup_queue on scheduler root
• mark_for_removal enqueues stream_id -> expires_at
• tick_closed_streams advances epoch and pops expired entries only

Validation

• dune runtest --no-buffer passes
• includes new test: closed stream cleanup does not depend on write polling