feat: harden runtime hooks and stream tool safety

[wxai-space]

Summary

• add max_tool_iterations as an explicit streaming tool-call loop limit while preserving the existing max_retry default behavior
• close stream tool-loop limit failures through _finish_run(...) so on_error, after_run, and run_end stay consistent
• bump the package version to 0.9.3 and update roadmap/runtime hook docs for the v0.9.3 hardening scope
• add regression coverage for hook failure isolation, model-error lifecycle closure, and stream tool-loop error payloads

Validation

• python -m compileall LightAgent
• python -m pytest -q tests/test_v065_core.py tests/test_v070_tracing.py tests/test_guardrails.py tests/test_memory_policy.py tests/test_lightflow.py tests/test_litellm_client.py tests/test_mcp_client_manager.py tests/test_shared_memory_pool.py tests/test_skill_manager_logging.py tests/test_vector_memory_adapter_example.py
• python -m pytest -q tests/test_v065_core.py tests/test_v070_tracing.py tests/test_guardrails.py tests/test_memory_policy.py
• git diff --check