Implement latency histograms to track request duration distribution and time-to-first-token (TTFB) for streaming requests

[psschwei]

Description:
Implement latency histograms to track request duration distribution and time-to-first-token (TTFB) for streaming requests.

Detailed Requirements:

1. Create histograms in mellea/telemetry/metrics.py:
   • mellea.llm.request.duration - Total request duration
   • mellea.llm.ttfb - Time to first byte/token (streaming only)
2. Define appropriate bucket boundaries:
   • Duration: [0.1, 0.25, 0.5, 1, 2.5, 5, 10, 30, 60, 120] seconds
   • TTFB: [0.05, 0.1, 0.25, 0.5, 1, 2, 5, 10] seconds
3. Add helper functions:
   • record_request_duration(duration_seconds, model, backend, streaming)
   • record_ttfb(ttfb_seconds, model, backend)
4. Integrate with backend instrumentation:
   • Measure total duration from request start to complete response
   • For streaming: measure time from request to first chunk
5. Add attributes:
   • gen_ai.system, gen_ai.request.model, mellea.backend
   • streaming - Boolean indicating streaming vs non-streaming

Files to Modify:

• mellea/telemetry/metrics.py - Add latency histograms
• mellea/telemetry/backend_instrumentation.py - Add timing helpers
• mellea/backends/openai.py - Record latency metrics
• Other backends as applicable

Metric Definitions:

Metric	Type	Unit	Attributes
mellea.llm.request.duration	Histogram	s	model, backend, system, streaming
mellea.llm.ttfb	Histogram	s	model, backend, system

Acceptance Criteria:

• Duration histogram captures all request latencies
• TTFB histogram captures streaming first-token time
• Bucket boundaries appropriate for LLM workloads
• Metrics have correct labels

[ajbozarth]

I should have a PR for this tomorrow, just need to finish reviewing and testing it