fix quic partial-frame desync via per-peer writer task

[fabracht]

Summary

• a timed-out send_to_peer used to cancel write_all mid-frame (it is not cancellation-safe), leaving a partial length-prefixed frame that the reused stream never resynced from (cluster: a timed-out send leaves a partial frame and desynchronizes the QUIC stream #143)
• each peer now has a writer task that solely owns its SendStream and writes one whole frame at a time with no timeout, so a frame is either written whole or the stream is torn down
• sends enqueue to a bounded per-peer channel via non-blocking try_send; a full queue drops the whole frame as SendQueueFull instead of truncating, and a slow peer can no longer block broadcast
• scope is framing safety only; peer-map removal and re-dial remain cluster: nodes never dial peers they discover, never retry failed dials, and never drop dead peers #140

Test plan

• cargo make clippy (pedantic, all feature combos incl. --no-default-features)
• cargo make test (1201 tests, 0 failures)
• new tests/quic_transport_framing.rs passes on the fix and fails on pre-fix code (verified by reverting only quic_transport.rs: assert!(!blocked) fires)

Follow-ups

out of scope here, tracked on the #140 mesh/reconnection track:

• the 1024-slot per-peer queue drops under sustained overload instead of the old blocking send; needs congestion-drop reconciliation
• a writer task that dies on a stream error leaves a stale peer-map entry with no outward re-dial