fix(cli): serialize LiteLlm graph models safely

[he-yufeng]

Fixes #5949.

The graph serializer currently passes model fields through when their value is not a simple collection or nested agent. For LiteLlm that leaves the runtime LiteLLMClient object inside the serialized graph payload, so JSON encoding the graph fails.

This patch serializes BaseLlm values through pydantic and drops the runtime llm_client field. The graph output keeps the public model configuration, for example {"model": "ollama_chat/llama3"}, while avoiding non-JSON-safe runtime clients.

Validation:

• .venv\Scripts\python.exe -m pytest tests\unittests\cli\utils\test_graph_serialization.py -q
• .venv\Scripts\python.exe -m py_compile src\google\adk\cli\utils\graph_serialization.py tests\unittests\cli\utils\test_graph_serialization.py
• .venv\Scripts\python.exe -m pyink --check src\google\adk\cli\utils\graph_serialization.py tests\unittests\cli\utils\test_graph_serialization.py
• .venv\Scripts\python.exe -c "import json; from google.adk.agents import LlmAgent; from google.adk.models.lite_llm import LiteLlm; from google.adk.cli.utils.graph_serialization import serialize_agent; agent=LlmAgent(name='repro', model=LiteLlm(model='ollama_chat/llama3')); data=serialize_agent(agent); print(data['model']); json.dumps(data); print('json ok')"
• git diff --check

Note: I used the repo-local virtualenv because the base Python environment still has an older google-genai that lacks types.AvatarConfig.