Performance metrics aggregation for workflow execution

[JaimeStill]

Migrated from tailored-agentic-units/tau-orchestrate#3

Summary

Collect and aggregate timing and resource metrics across workflow execution.

Requirements

• Node execution duration (start/complete timestamps already in events)
• Workflow total duration
• Parallel worker utilization (active workers, queue depth)
• Per-step latency distribution across chain executions
• Error rates and categorization
• Aggregation into summary statistics (min, max, mean, p50, p95, p99)

Design Constraints

• Zero-overhead default: advanced features are opt-in; NoOpObserver behavior must remain zero-cost
• Registry compatible: new observers must integrate with the existing string-based registry pattern
• No core package coupling: consider a separate package (e.g., orchestrate/metrics) to avoid adding dependencies to core orchestrate/observability

Acceptance Criteria

• Performance metrics are aggregatable into summary statistics without external tooling
• All new observer implementations pass through the existing registry pattern
• No performance regression for existing NoOpObserver and SlogObserver usage