split quic peer sends into control and bulk priority lanes

[fabracht]

Summary

• under sustained bulk traffic the single 1024-slot per-peer queue could drop or delay heartbeats and raft messages behind bulk data, risking false death-detection and election churn (cluster: congestion-drop reconciliation and stale-peer redial after writer-task death #146)
• each peer now has two lanes: a small control channel (heartbeats, raft, death/drain notices, partition updates) and the existing bulk channel; the writer drains control with priority
• a full bulk queue no longer drops control-plane sends
• addresses the congestion-drop half of cluster: congestion-drop reconciliation and stale-peer redial after writer-task death #146; stale-peer redial is still tracked there

Notes

• control and bulk share one QUIC stream, so a single in-flight bulk write_all can still delay a control frame by one frame; fully removing that needs a separate stream per lane, out of scope here
• bulk frames still queued at teardown/replacement may be dropped; recoverable via heartbeat/raft retry and replication catchup

Test plan

• cargo make clippy (pedantic, all feature combos)
• cargo make test (1202 tests, 0 failures)
• new control_plane_survives_a_full_bulk_queue test fails on the pre-change single-lane code and passes after; framing test still passes