fix: delete per-session prometheus gauge series on Syncer.Close

[mxssl]

Problem

The gauges chotki_sync_sessions (labels id, kind, version), chotki_sync_opened_snapshots and chotki_sync_opened_iterators (labels id, version) use the connection name as the id label. Listen-side names (listen:<uuid>:<ip>:<port>) are unique per connection, and the series were never removed when a session terminated — they accumulate until the process restarts.

Observed on a long-running production deployment (2026-07-05): 84,320 stale chotki_sync_sessions series in state 4 (SendNone) vs 135 live sessions; long-lived nodes exposed ~46k samples per /metrics scrape vs ~10k on freshly-restarted ones. Unbounded label cardinality like this eventually pushes /metrics past scraper response-size limits, at which point the entire scrape gets dropped silently.

Fix

Syncer.Close() now deletes the per-session series. network.Peer.Close() calls it only after wg.Wait() has joined the read/write loops, so no Feed/Drain runs concurrently with the deletion (and a straggler double-Close is self-cleaning because the deletes are deferred to function exit).

One subtlety: alerting on leaked pebble resources relies on opened_snapshots/opened_iterators staying nonzero when a close fails. Deleting those series unconditionally would blind it. So:

• chotki_sync_sessions (feed + drain) series are always deleted, including the nil-Host early-return path.
• Snapshot/iterator series are deleted only when every close succeeded. Mid-session close failures in Feed() (SendDiff/SendEOF paths) and FeedDiffVV() nil the handles while leaving the gauge at 1, so new snapCloseFailed/iterCloseFailed flags (set under the lock) tell Close() to keep those series as the leak signal.

Testing

New white-box tests in replication/sync_metrics_test.go: clean close removes all per-session series; close with a failing snapshot keeps the snapshot series; mid-session close failure (flags set) keeps snapshot+iterator series; nil-Host close still removes state series. Full go test ./... passes.

Review notes

Both reviewers noted that Feed/FeedDiffVV mutate snap/ffit/vvit outside sync.lock while Close() reads them under it — that is pre-existing behavior, safe under the current Peer.Close ordering guarantee, and left untouched to keep this change minimal. Related follow-up candidate (not in this PR): chotki_outbound_packet_count (CounterVec, per-peer name label) leaks series the same way.

[mxssl]

Closing this — the per-session prometheus gauge cleanup on Syncer.Close is already fixed in #21.