fix: make full nodes not resync infinitely on flush

[pablocampogo]

Fix full-node stall: trigger an active resync when the block inbox backs up

Summary

Full nodes (non-validators) could permanently stall on new block messages and only recover after a restart. The existing 80% inbox dead-letter flush did not fix it — in one production incident the flush fired 16 times over ~2 hours and the node still never resynced. This PR makes a backed-up block inbox trigger an active resync instead of just dropping the backlog.

Root cause

Inbound BLOCK messages funnel into a single shared channel consumed by either ListenForBlock() (steady state) or Sync() (catch-up). The only pre-existing way for a synced full node to re-enter Sync() was the syncDetector (NewHeightTracker) in ListenForBlock(), which requires ≥ 1/3 of peers to each deliver a fresh ErrNewHeight block.

That recovery path structurally starves exactly when it's needed:

• When a full node falls behind, the shared block inbox saturates. The non-blocking producer drops the newest messages, and the DLQ drops the entire backlog at 80%.
• Those drops discard the very ErrNewHeight gossip that syncDetector needs to count toward its 1/3 threshold, so the trigger never fires.
• Gossiped blocks are de-duplicated (MessageCache) and applied strictly in-order, so any height dropped by the flush never arrives again on its own. Only Sync() re-requests specific heights — and there was no path from "I dropped blocks" → "request them back."

Net effect: the node churns (flush → refill → flush) forever and never transitions to sync. A restart worked only because boot unconditionally runs Sync(). This was full-node-only because the DLQ is disabled for validators.

Additionally, the flush was actively harmful during a sync: it would drop the in-order block responses Sync() depends on, causing timeout/re-request livelock.

Changes

• p2p/p2p.go: added a syncing flag (SetSyncing/IsSyncing) and a coalesced mustResync channel exposed via MustResync() / signalResync().
• p2p/conn.go (dropInboxIfBackedUp): for the BLOCK topic on a full node, do nothing while a sync is already running (so Sync() keeps its in-order responses, matching validator behavior), otherwise signal a resync before draining the backlog.
• controller/block.go (ListenForBlock): now selects on both Inbox(Block) and MustResync(); on a resync signal it launches Sync() and cleanly hands off consumption.
• controller/consensus.go: Sync() is idempotent via isSyncing.CompareAndSwap (prevents duplicate Inbox(Block) consumers now that there are multiple trigger paths) and toggles P2P.SetSyncing(true/false) in Sync()/finishSyncing().

Behavior after fix

• A full node that falls behind now switches to an active resync that re-requests missing heights in order, rather than relying on the peer-count heuristic that the backup itself was starving.
• The DLQ flush is suppressed during sync, so full-node sync behaves like validator sync (no dropped responses).
• Validators remain fully exempt from the DLQ.