feat(worker): add priority-based consolidation bank scheduling

[nicoloboschi]

Summary

Closes #1715

• Adds HINDSIGHT_API_WORKER_CONSOLIDATION_BANK_PRIORITY env var to control which banks' consolidation tasks are claimed first when a slot opens
• Prevents large banks (e.g., 486K-node shadow-meetings) from being starved by many small banks cycling through limited global consolidation slots
• Uses tiered claiming — each priority level is a separate index-friendly query (no JOINs or computed ORDER BY). Number of queries = number of distinct priority tiers (typically 2). Zero overhead when unset
• Pattern wildcards supported (e.g., shadow-*:10,staging-*:5,*:1), using LIKE ANY / NOT LIKE ALL in PostgreSQL, expanded to OR/AND clauses for Oracle
• Bank serialization (max 1 concurrent consolidation per bank) is preserved regardless of priority
• Backwards compatible: when unset, behavior is identical to current ORDER BY created_at

Design decision: priority queues vs per-bank slot reservation

The original issue proposed per-bank slot allocation (SLOTS_PER_BANK). We chose priority-based scheduling instead because:

• Consolidation batches are short-lived (capped at 100 memories, then reschedule), so the real problem is scheduling order, not slot reservation
• Priority queues waste no capacity — when a high-priority bank has no pending work, other banks use the slots freely
• This matches standard distributed task system patterns (Celery, SQS, Sidekiq)

Test plan

• 8 unit tests for config parsing (TestParseBankPriority)
• 5 integration tests against real PostgreSQL (TestConsolidationBankPriority):
  • High-priority bank claimed before low-priority despite newer created_at
  • Wildcard pattern matching (shadow-* matches shadow-meetings, shadow-people)
  • Catch-all default for unlisted banks
  • Backwards compat when priority unset
  • Priority respects bank serialization (busy high-priority bank still excluded)
• All 80 existing worker tests pass with no regressions
• Lint passes