Epic: Comprehensive Metrics System

[psschwei]

Description:
Add metrics collection for key operational indicators: token usage, latency, costs, errors, and sampling/validation outcomes. This provides quantitative observability complementing the existing tracing system.

Current State:

• No metrics collection system exists
• Token usage is extracted during backend operations but only used for tracing attributes
• No latency histograms, error counters, or cost tracking
• No Prometheus or OTLP metrics export

Acceptance Criteria:

• Token usage counters by model/backend
• Latency histograms including TTFB for streaming
• Cost tracking with pricing registry
• Error counters by semantic type
• Prometheus and OTLP export options
• Operational counters for sampling/validation/tools