Add OpenAI-shaped response trace models

[soon-dubu]

Summary

• add OpenAI-shaped Trace and generic Span[TSpanData] response models without changing runtime tracing
• define discriminated workflow, GUI-step, control-flow, agent, and agent-action span-data models
• support structural child spans with explicit null timestamps
• use starting_url for agent actions so it matches GUI-step URL semantics
• add workflow chat_input and effective input_variables, where explicit inputs override authored defaults
• require GUI agent-step inputs to report the effective reasoning_mode
• use user-facing GUI-step names and omit the ambiguous GUI-step description field
• keep group_id as an optional caller-supplied cross-trace correlation identifier that is distinct from per-run request IDs
• expose the response models from the local narada SDK; narada-pyodide is intentionally unchanged
• centralize the tracing public surface and enforce taxonomy, statuses, validation, hierarchy, and exports with tests

Scope

This is Stage 1 only: typed response models, local SDK public exports, serialization behavior, and tests. The supported sources are GUI Agent Studio workflow traces and local Python SDK responses.

This PR does not add AgentResponse.trace, conversion adapters, trace collection, bundle validation, runtime producer changes, billing attribution, or Core/Productivity agent-action instrumentation.

Existing action_trace and workflow_trace response behavior is unchanged.

Testing

• focused shape and public-export suites: 31 passed
• Ruff check and format passed

Codex sibling refs

• caddie:
• frontend: https://github.com/NaradaAI/frontend/pull/2001
• architecture-docs: https://github.com/NaradaAI/architecture-docs/pull/74
• desktop-automation-app:
• api-docs:

[soon-dubu]

Can remove the per-run request identifier portion here