feat(integrations): add phase 3 agent framework adapters

[sinewaveai]

What

Adds Phase 3 agent-framework integrations from MCP_RUNTIME_INTEGRATIONS_PRD.md:

• OpenAI Agents SDK: prooflayer.integrations.openai_agents.ProofLayerGuardrail
• CrewAI: prooflayer.integrations.crewai.SecurityMiddleware
• AutoGen: prooflayer.integrations.autogen.SecurityMiddleware
• Semantic Kernel: prooflayer.integrations.semantic_kernel.SecurityMiddleware
• Pydantic AI: prooflayer.integrations.pydantic_ai.SecurityMiddleware

Also adds shared agent security primitives for agent input/output scanning, unsafe handoff detection, role drift heuristics, cross-agent instruction smuggling heuristics, callable tool wrapping, optional package extras, docs, examples, and tests.

Why

Phase 1 established common runtime primitives and Phase 2 added LangChain MCP + LlamaIndex wrappers. Phase 3 extends the same security contract to agent-framework runtimes so ProofLayer can inspect agent inputs, outputs, tool calls, and handoff/delegation messages without changing each framework's native invocation flow.

Verification

• Fresh Python 3.12 editable install: pip install -e .
• Example scripts ran successfully for OpenAI Agents, CrewAI, AutoGen, Semantic Kernel, and Pydantic AI
• Targeted integrations: 93 passed, 1 warning
• Full suite: 393 passed, 1 warning, 19 subtests passed
• Phase 3 coverage: 81.28% over new adapter modules
• Formatting: black --check passed
• Type checking: mypy prooflayer/ --ignore-missing-imports passed

Notes

Optional framework dependencies stay out of the base import path. The adapters use dependency-light proxies and fake-runtime tests so base CI remains stable while customers can install framework-specific extras when needed.