cluster: congestion-drop reconciliation and stale-peer redial after writer-task death

[fabracht]

Follow-ups from #143 / #145 (per-peer QUIC writer task), out of scope there. Both belong to the #140 full-peer-mesh / reconnection track.

• congestion-drop reconciliation: send_to_peer now drops a whole frame with SendQueueFull once the 1024-slot per-peer queue fills, instead of the old blocking send. this removes event-loop head-of-line blocking but changes overload semantics — under a bursty-but-recoverable stall the old path eventually delivered with latency, the new path drops once the queue backs up. needs a reconciliation / adaptive-queue story so cluster traffic dropped under congestion is reconciled.
• stale-peer redial after writer-task death: when a peer_writer_task exits on a stream write error, its peer-map entry is left in place and nothing re-dials outward, so later sends return SendFailed/Disconnected until an incoming connection replaces the entry. same reconnection gap as cluster: nodes never dial peers they discover, never retry failed dials, and never drop dead peers #140.

neither affects framing safety, which is fixed in #145.