feat(agent): forward model settings and skip forced tool_choice under Claude thinking [PC-4672]

[tudormatei1]

Two related changes for model-specific settings.

chat_model_factory.get_chat_model gets a model_settings param that's forwarded to the new-client factory, so the provider-native settings.modelSettings from agent.json reaches the actual LLM call. Defaults to None, so the legacy client path is unaffected.

The ReAct graph gets a reasoning_enabled flag on AgentGraphConfig, threaded into the LLM node and router. When it's set, the graph stops forcing tool_choice="any" and stops raising THINKING_LIMIT_EXCEEDED on tool-less turns. This is for Claude with extended thinking: the Anthropic API rejects a forced tool_choice while thinking is on (400), and Claude's thinking blocks are legitimate tool-less turns that shouldn't trip the loop guard. The run is still bounded by llm_messages_limit. The flag is off by default, so every other model keeps the normal force-and-terminate behavior — importantly OpenAI, which allows forced tool_choice alongside reasoning and would otherwise loop.

Bumps the uipath-langchain-client pins to >=1.18.0,<1.19.0 (the model_settings param on get_chat_model only exists there) and the package to 0.14.18. Core uipath-llm-client is unchanged. The uv.lock regen is deferred until uipath-langchain-client 1.18.0 is on PyPI