Improve consistency and DRY across Penelope targets and SDK telemetry integrations

[harry-rhesis]

Summary

Address cross-cutting consistency gaps and DRY opportunities across the Penelope target implementations and SDK telemetry integrations, identified during the review of PR #2057.

Background / Context

PR #2057 (Pydantic AI integration) revealed several inconsistencies and duplication patterns that exist across all targets and integrations, not just Pydantic AI. These are best addressed in a separate PR to keep #2057 focused.

Deliverables

• Add try/except ImportError guards in penelope/src/rhesis/penelope/targets/__init__.py so that missing optional dependencies (langchain, langgraph, pydantic-ai) degrade gracefully instead of breaking the entire import
• Add native a_send_message() to LangChainTarget and LangGraphTarget (using ainvoke / arun) instead of relying on the base class thread-pool fallback, matching PydanticAITarget's pattern
• Extract PatchState base class or factory from the duplicated AgentPatchState (pydantic_ai.py) and GraphPatchState (langgraph.py) into integrations/base.py
• Add LLM-level child spans (ai.llm.invoke) inside Pydantic AI agent runs, matching the visibility LangChain/LangGraph provide for individual model calls within an agent run

Acceptance Criteria

• import rhesis.penelope.targets succeeds on a minimal install without langchain, langgraph, or pydantic-ai
• LangChainTarget and LangGraphTarget have native async paths that do not use asyncio.to_thread
• GraphPatchState and AgentPatchState share a common base, reducing code duplication
• Pydantic AI agent runs produce ai.llm.invoke child spans for individual model calls

Additional Context

Related to PR #2057: #2057

[rhiannalitchfield]

Opened #2115 to close this out. Accounting for the original deliverables, since two were reshaped by the #2104 translator migration:

• Import guards: already satisfied by the deferred-import pattern from Add Pydantic AI integration (auto-instrumentation + Penelope target) #2057; the PR extends the regression test to also cover agent_framework/MAFTarget.
• Native async for LangChainTarget/LangGraphTarget: delivered in the PR via native ainvoke() plus an async file-attachment path.
• PatchState base extraction: moot. AgentPatchState was deleted when Migrate Pydantic AI integration to native instrumentation translator #2104 replaced the Agent.run monkey-patch, leaving GraphPatchState (langgraph) as the only patch-state class, so there is no duplication left to extract.
• ai.llm.invoke child spans for Pydantic AI: delivered by Migrate Pydantic AI integration to native instrumentation translator #2104's native per-model-call chat spans.

The PR also folds in the follow-ups accumulated during the #2104 review: the shared AncestryRegistry consolidation into genai.py, and porting the dedup-flag restore and fail-closed enable() hardening to the MAF integration.