fix: remove global add_function_to_prompt — breaks native tool calling

Setting `litellm.add_function_to_prompt = True` globally forces ALL
models through text-based tool calling, even models that support
native function calling (Groq, OpenAI, Anthropic).

When this flag is set, LiteLLM injects tool definitions into the
system prompt as text. Models then output XML-style function tags
(`<function=name {...} </function>`) instead of proper `tool_calls`
JSON. Providers like Groq reject this with `tool_use_failed`.

Proof: Direct `litellm.completion()` without this flag returns proper
`tool_calls` JSON with `finish_reason: "tool_calls"`. With the flag,
the same model fails.

The fix removes the global default. Models that need text-based tool
calling can opt in per-instance:

    LiteLlm(model="ollama/qwen2", add_function_to_prompt=True)

Models with native tool calling work without any flag:

    LiteLlm(model="groq/llama-3.3-70b-versatile")

Fixes: kagent-dev/kagent#1532
Related: huggingface/smolagents#1119, BerriAI/litellm#11001

Author: vitas

src/google/adk/models/lite_llm.py

@@ -185,7 +185,12 @@ def _ensure_litellm_imported() -> None:
 
   import litellm as litellm_module
 
-  litellm_module.add_function_to_prompt = True
+  # Do NOT set litellm.add_function_to_prompt = True globally.
+  # That flag injects tool definitions into the system prompt as text,
+  # which breaks models that support native tool calling (Groq, OpenAI,
+  # Anthropic). Those models then output XML-style function tags instead
+  # of proper tool_calls JSON, causing "tool_use_failed" errors.
+  # See: https://github.com/google/adk-python/issues/XXXX
 
   globals()["litellm"] = litellm_module
   for symbol in _LITELLM_GLOBAL_SYMBOLS: