connection: add per-peer exponential reconnection backoff

[Jolah1]

Closes #918.

Summary

Today the reconnection loop retries every disconnected peer on a fixed PEER_RECONNECTION_INTERVAL (60s). For a peer that is persistently unreachable this means a reconnect attempt every minute indefinitely, which is wasteful and noisy. This adds per-peer exponential backoff so repeated failures are spaced out, while a successful connection (or a user-initiated connect) resets a peer back to the base interval.

What this does

• Introduces PeerReconnectState in connection.rs, tracked per peer in a Mutex<HashMap<PublicKey, PeerReconnectState>> on ConnectionManager.
• Each consecutive failure schedules the next retry at the current backoff and doubles it, capped at PEER_RECONNECTION_MAX_INTERVAL (30 min).
• The reconnect loop gates each peer on is_reconnect_due rather than attempting every peer every tick.
• Backoff state is cleared on a successful connect (do_connect_peer), on explicit disconnect(), and when a peer is removed after its last channel closes.
• Already-connected peers have their state cleared each tick, so an inbound
  reconnection also resets backoff.

Tests

• reconnect_state_doubles_until_capped - verifies doubling and the cap.
• reconnect_state_schedules_relative_to_failure_time - verifies the next retry is scheduled relative to the failure instant.

A note on "configurable"

The issue title says "configurable." This PR keeps the base/cap as pub(crate) constants (matching the existing reconnection-interval constant) rather than exposing them on Config. Happy to wire them into Config as a follow-up (or here) if reviewers prefer — wanted to keep the surface minimal first.

Dependency / draft status

Marking this as draft: it builds on #895, which relocates last-channel
peer removal into the event.rs ChannelClosed handler. Once #895 lands I
will rebase and move the clear_reconnect_state call to sit alongside the
relocated removal (and pick up the now-async remove_peer .awaits from #919). Will un-draft after #895 is merged.

[ldk-reviews-bot]

👋 Hi! This PR is now in draft status.
I'll wait to assign reviewers until you mark it as ready for review.
Just convert it out of draft status when you're ready for review!

[Jolah1]

Un-drafting now that #895 has landed. Changes since the initial draft:

Rebased on main. As promised above, the backoff clear that lived in
close_channel_internal moved to #895's relocated peer removal in the
ChannelClosed event handler, so backoff state is dropped exactly where the
peer leaves the store.

After another self-review pass:

• Backoff bookkeeping is now anchored to the reconnection loop's tick instant
  instead of each attempt's completion time. Previously a first failure was
  scheduled at attempt_end + 60s, which always slipped past the next 60s
  tick and effectively doubled the first retry delay.
• The loop now skips a peer that already has a connection attempt in flight
  (has_pending_connection) rather than subscribing to it. This avoids
  recording someone else's failure — possibly for a different, user-supplied
  address — as a failure of the persisted address.
• Backoff state is pruned against the persisted peer store each tick. This
  guards a race where a failed in-flight attempt could resurrect state for a
  peer that was concurrently removed (leaking the entry), and makes cleanup
  self-healing if future code adds remove_peer call sites.
• connect_peer_if_necessary clears backoff on its already-connected early
  return, since the peer is demonstrably reachable.
• Simplified record_reconnect_attempt(&Result) into
  record_reconnect_failure called only on Err; do_connect_peer is now
  the single owner of clear-on-success.

One known limitation, documented in a comment: a peer that connects inbound
and drops again entirely within one tick keeps its accumulated backoff, since
connectivity is only observed at tick time. Clearing eagerly on inbound
connects would require hooking the LDK peer_connected handlers, which felt
out of scope here .

cc @tnull