Defer LSPS4 HTLC forwarding until channel usable

[amackillop]

LDK fires peer_connected after the TCP+Init handshake but before
channel_reestablish completes. During this window, channels exist
but are not yet usable (is_usable=false). The previous code
forwarded HTLCs unconditionally in peer_connected and
htlc_intercepted, causing ~10% of payments to fail on reconnect.

Distinguish two connected-peer cases in htlc_intercepted and
peer_connected:

• Channels exist but none usable (reestablish in progress): defer
  the HTLC for timer-based retry via process_pending_htlcs().
• No channels exist (first payment, JIT open needed): proceed
  immediately so calculate_htlc_actions_for_peer can emit the
  OpenChannel event.

process_pending_htlcs (5s timer) only retries the reestablish
case (channels exist, waiting to become usable). It must not
handle the no-channel case to avoid emitting duplicate
OpenChannel events while a JIT open is already in flight.

Remove the re-check race in htlc_intercepted that could fire
concurrently with peer_connected. The webhook + peer_connected
path is the single owner of the offline-peer reconnect flow.

Blocking inside peer_connected was also considered but rejected:
there is no LDK event for "channel usable after reconnect" to
wake on, so it would require a spin-wait with arbitrary timeout.
A timer-based retry is cleaner and avoids holding the lock.

Summary:

Path	Usable channels	Channels reestablishing	No channels (JIT)
htlc_intercepted (connected)	Forward	Defer	Trigger OpenChannel
htlc_intercepted (offline)	—	—	Store + webhook
peer_connected	Forward	Defer	Trigger OpenChannel
process_pending_htlcs (timer)	Forward	Wait	Skip (no duplicate)
channel_ready	Forward	—	—