Implement error counters categorized by semantic error type for better operational visibility #465

Description:
Implement error counters categorized by semantic error type for better operational visibility.

Detailed Requirements:

1. Create error type enum/constants in mellea/telemetry/metrics.py:
   • rate_limit - Rate limiting / 429 errors
   • timeout - Request timeouts
   • content_policy - Content policy violations
   • auth - Authentication/authorization errors
   • invalid_request - Malformed request errors
   • transport_error - Network/connection errors
   • server_error - Provider server errors (5xx)
   • unknown - Uncategorized errors
2. Create counter:
   • mellea.llm.errors - Error count by type
3. Create error classifier function:
   • classify_error(exception) -> str - Map exceptions to semantic types
   • Handle provider-specific exception types (OpenAI, etc.)
4. Create helper:
   • record_error(error_type, model, backend, exception_class)
5. Integrate with backend error handling

Files to Modify:

• mellea/telemetry/metrics.py - Add error counter and classifier
• mellea/telemetry/backend_instrumentation.py - Add error recording
• mellea/backends/openai.py - Record errors on exceptions
• Other backends as applicable

Metric Definition:

Error Type Mapping (examples):

Acceptance Criteria:

• All errors classified into semantic types
• Counter increments with correct labels
• Provider-specific exceptions handled
• Unknown errors categorized as "unknown"