Create a standardized error classification system that maps exceptions to semantic error types for consistent error reporting in traces

[psschwei]

Description:
Create a standardized error classification system that maps exceptions to semantic error types for consistent error reporting in traces.

Detailed Requirements:

1. Define semantic error types in mellea/telemetry/errors.py:
   • rate_limit - Rate limiting errors
   • timeout - Request timeouts
   • content_policy - Content policy violations
   • auth - Authentication errors
   • invalid_request - Bad request errors
   • transport_error - Network errors
   • server_error - Provider server errors
   • validation_error - Mellea validation failures
   • tool_error - Tool execution errors
2. Create exception classifier:
   • classify_exception(exception) -> ErrorType
   • Handle provider-specific exceptions (OpenAI, Anthropic, etc.)
3. Create helper for setting span status:
   • set_semantic_error(span, exception) - Sets status and error attributes
4. Integrate with existing set_span_error() function
5. Add error.type span attribute with semantic type

Files to Create:

• mellea/telemetry/errors.py

Files to Modify:

• mellea/telemetry/__init__.py - Export error utilities
• mellea/telemetry/backend_instrumentation.py - Use semantic errors

Error Type Attributes:

Attribute	Description
error.type	Semantic error category
error.message	Exception message
exception.type	Python exception class name

Acceptance Criteria:

• All common exceptions mapped to semantic types
• Span status correctly set on errors
• Error type attribute added to spans
• Provider-specific exceptions handled

[ajbozarth]

Closing as part of the tracing epic (#444) re-plan. This sub-issue predates the plugin/hooks pattern (PR #653) and the GenerationMetadata partition (PR #908); its spec has drifted.

The metrics epic (issue #465) already built a matching error taxonomy (classify_error + 8 ERROR_TYPE_* constants in mellea/telemetry/metrics_plugins.py), so rather than create a parallel module the refreshed plan reuses it.

Replacement: split across #1053 (share metrics' classify_error with tracing plugins) and #1047 (wire the semantic error category onto spans via a generation_error plugin hook).

See #444 for the full refreshed plan.