Support evaluation of models with interleaved thinking

[lucenzhong]

To support evaluation of models with interleaved thinking, the input message should preserve reasoning_content from previous turns to maintain reasoning consistency.

The relevant code changes are as follows:

1. Enable thinking when calling the model
2. Support returning reasoning_content in llm.py and schema.py

mcp-atlas/services/mcp_eval/mcp_completion/llm.py

Lines 75 to 79 in 867003a

	assistant_message = AssistantMessage(
	role="assistant",
	content=response.choices[0].message.content,
	tool_calls=tool_calls,
	)

mcp-atlas/services/mcp_eval/mcp_completion/schema.py

Lines 36 to 41 in 867003a

	class AssistantMessage(BaseModel):
	"""Assistant message."""
	role: Literal["assistant"]
	content: Optional[str] = None
	tool_calls: Optional[List[ToolCall]] = None

Modified to:

 assistant_message = AssistantMessage( 
     role="assistant", 
     content=response.choices[0].message.content, 
     tool_calls=tool_calls, 
     reasoning_content=response.choices[0].message.reasoning_content,
 ) 

class AssistantMessage(BaseModel):
    """Assistant message."""

    role: Literal["assistant"]
    content: Optional[str] = None
    tool_calls: Optional[List[ToolCall]] = None
    reasoning_content: Optional[str] = None