fix: fenced lock release, wire retry processor, guard indexer metrics on RPC fail, yield tests

[dikachim060-wq]

Summary

• [Backend] defaultChecker releaseLock() deletes the Redis lock unconditionally, letting a long run delete another instance's lock #1163 — Added deleteIfMatch(key, expectedValue) to cacheService.ts and updated defaultChecker.acquireLock to store the fencing token in this.currentLockValue. releaseLock now calls deleteIfMatch instead of unconditional delete, so a run that outlives the 10-minute TTL cannot delete a lock acquired by a second instance.
• [Backend] startWebhookRetryProcessor is dead and the wired scheduler uses a different backoff than webhookService stores #1167 — Removed the dead webhookRetryScheduler.ts (used a different backoff, ignored next_retry_at). index.ts now wires startWebhookRetryProcessor / stopWebhookRetryProcessor instead. Added a doc comment to webhookRetryProcessor.ts describing the effective 5m/15m/45m backoff schedule and 4-attempt cap.
• [Testing] yieldHistoryService withdraw cost-basis and XDR share-decoding paths are uncovered #1171 — Extended yieldHistoryService.test.ts with four new tests: Withdraw reduces cost basis proportionally, EmergencyWithdraw follows the same path, base64 XDR value is decoded (BigInt shares extracted), and malformed XDR falls back gracefully without throwing.
• [Backend] RPC failure to fetch chain tip records indexer_lag_ledgers=0, masking outages and defeating the lag alert #1201 — In eventIndexer.pollOnce, skip the recordIndexerLedgers call when latestLedger === 0 (RPC failure). Gauges are left at their last known values so the behind-chain-tip alert fires correctly during an outage instead of silently reading 0.

Test plan

• npm test -- --testPathPattern yieldHistoryService passes including new Withdraw/XDR tests
• Confirm webhookRetryScheduler.ts is removed and no import references remain
• During a simulated RPC failure indexerLagLedgersGauge stays at its last non-zero value

closes #1163
closes #1167
closes #1171
closes #1201