fix: bound root-chain calls to prevent node-wide deadlock

[pablocampogo]

fix: bound root-chain calls to prevent node-wide deadlock

Full nodes intermittently wedged until restart: the BLOCK and
BLOCK_REQUEST inboxes filled to capacity and dropped every message
("CRITICAL: Inbox BLOCK queue full"). Both inbox consumers require the
controller mutex, which was being held forever by block processing
stuck in an unbounded root-chain call.

Only full nodes were affected because validators stay on the in-memory
committee cache: their cached root-chain info tracks the live root
height, so LoadCommittee resolves from sub.Info without a network call.
Full nodes repeatedly fall behind and re-sync, crossing checkpoint
heights and certificates with stale root heights, which forces
committee/checkpoint lookups onto the remote HTTP path -- and that path
had no timeout, so one request stuck on a half-open connection (NAT
drop, root-chain RPC restart) held the controller mutex forever.

Two fixes:

• Use an http.Client with a 10s timeout for RCSubscription's on-demand
  root-chain calls (LoadCommittee, GetCheckpoint, lottery, dex batch,
  etc.). A hung request now surfaces as an error and the block is
  simply re-requested instead of wedging block processing.

• Add a read deadline + ping-refreshed liveness check to the root-chain
  subscription websocket. A silently dead connection is now detected
  and redialed instead of hanging Listen() forever, which left the
  cached root-chain info stale and forced lookups onto the broken
  remote path.