feat(#223): map logical usageMetrics names to catalog metrics so models price from the catalog

[elecnix]

Summary

Closes #223. Connects the pricing catalog to DAG models. Previously the catalog (live Infracost + seed) was unreachable from real models: nodes author usageMetrics with logical names (natHours, keysCount), but the catalog is keyed by usage_metric names (NAT-Gateway-Hour, KMS-Key-Month). The engine queried the catalog with the raw logical key → always missed → silently fell back to embedded pricingRates, contradicting Principle 13.

Changes

• Per-handler catalog_metrics map (logical → catalog usage_metric) on the resource types that correspond to catalog rows: NAT Gateway, VPC Endpoint, Elastic IP, CloudWatch Logs + Metrics/Alarms, KMS, Secrets Manager, ECR, Route53, Data Transfer. Derived directly from each handler's _*_cost function so the names are exact.
  • Keyed per handler, not per service — resources of the same service reuse a logical name for a different catalog metric: dataProcessedGb → NAT-Gateway-DataProcessed vs VPC-Endpoint-DataProcessed; storedGb → ECR-Storage vs CloudWatch-Log-Storage.
• ResourceRegistry.resolve_catalog_metric(address, logical) resolves via the handler that owns the node's resourceAddress.
• Engine (flat + tiered cost paths) retries the catalog query with the mapped name when the raw name misses, before the pricingRates fallback.

Backward compatibility

Fully preserved: the raw metric name is tried first (nodes already using catalog names keep working), and unmapped logical names / non-handler addresses fall back to pricingRates exactly as before.

Test plan

• resolve_catalog_metric: per-handler resolution incl. shared-logical-name disambiguation (NAT vs VPC dataProcessedGb), unknown address / unmapped name → None.
• End-to-end CostAggregator: a NAT node authored with natHours prices from the catalog ($32.85 = 730 × $0.045), reachable only if the mapping resolved.
• Unmapped-metric → pricingRates fallback still applies.
• Full suite: 908 passed locally.

Note: handlers whose example models use ad-hoc metric names not equal to their valid_metrics (e.g. ALB lcus, Lambda gb_seconds) are not mapped here — those names are ambiguous and can be aligned/added incrementally. This PR establishes the mechanism and covers the handlers with unambiguous catalog metrics.