Trace latency: add metrics summary fast path

[thorrester]

Parent: #281

Goal

Add a fast path for dashboard metrics and summary queries that should not require repeatedly scanning trace-span Parquet files.

Dashboards tend to issue repeated aggregate queries over recent windows. They should use precomputed or narrower summary data where correctness allows it.

Scope

• Identify dashboard metrics and summary requests that can be served from summary tables or bounded aggregate paths.
• Keep raw trace-span scans available for requests that need span-level detail.
• Add freshness behavior that makes summary results safe under ongoing ingestion.
• Track fast-path hit rate, fallback rate, latency, and object-store request count.

High-level design

The goal is not to move the whole analytical workload to Postgres. The goal is to avoid using the most expensive raw-span path for queries whose result shape is already summary-oriented.

This issue should clarify which queries are eligible for the fast path and which must stay on the raw DataFusion path.

Acceptance criteria

• Eligible dashboard queries use the metrics/summary fast path.
• Ineligible queries fall back to the existing raw-span DataFusion path.
• Freshness behavior is explicit and tested.
• Metrics show fast-path hit rate and fallback reasons.
• Benchmarks show dashboard latency and object-store request-count impact.