Issue Description
Deploying an ADK agent with MCPToolset to Vertex AI Agent Engine fails with TypeError: cannot pickle 'TextIOWrapper' instances. The agent works perfectly in local development but cannot be deployed due to serialization issues during the deepcopy process.

Steps to reproduce the behavior:

1. Create an ADK agent with MCPToolset using SseServerParams
2. Wrap the agent with reasoning_engines.AdkApp
3. Deploy to Agent Engine using agent_engines.create()
4. See error: TypeError: cannot pickle 'TextIOWrapper' instances

Code to reproduce:

from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset
from google.adk.tools.mcp_tool.mcp_session_manager import SseServerParams
from google.adk.agents import LlmAgent
import vertexai
from vertexai import agent_engines
from vertexai.preview import reasoning_engines

PROJECT_ID = "your-project-id"
LOCATION = "your-deployment-location"
STAGING_BUCKET = "gs://your-staging-bucket"

vertexai.init(
    project=PROJECT_ID,
    location=LOCATION,
    staging_bucket=STAGING_BUCKET,
)

root_agent = LlmAgent(
    name="WikipediaQueryAgent",
    model="gemini-2.0-flash",
    description="An AI agent that can query Wikipedia",
    instruction="You are a helpful agent who can answer user questions.",
    tools=[
        MCPToolset(
            connection_params=SseServerParams(
                url="<your-MCP-Server-URL>/sse"
            )
        )
    ]   
)

app = reasoning_engines.AdkApp(
    agent=root_agent,
    enable_tracing=False,
)

# This line fails with serialization error
remote_app = agent_engines.create(
    agent_engine=app,
    requirements=[
        "google-cloud-aiplatform[adk,agent_engines]",
        "google-adk[mcp]"
    ]
)

Expected behavior
The agent should deploy successfully to Agent Engine. MCPToolset is documented as a supported tool type for ADK agents, so deployment should work without serialization errors.

Screenshots
Error traceback:

Deploying to Agent Engine...
Traceback (most recent call last):
  File "agent1.py", line 47, in <module>
    remote_app = agent_engines.create(
                 ^^^^^^^^^^^^^^^^^^^^^
  File ".../vertexai/agent_engines/__init__.py", line 148, in create
    return AgentEngine.create(
           ^^^^^^^^^^^^^^^^^^^
  File ".../vertexai/agent_engines/_agent_engines.py", line 422, in create
    agent_engine = _validate_agent_engine_or_raise(agent_engine)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../vertexai/agent_engines/_agent_engines.py", line 863, in _validate_agent_engine_or_raise
    agent_engine = agent_engine.clone()
                   ^^^^^^^^^^^^^^^^^^^^
  File ".../vertexai/preview/reasoning_engines/templates/adk.py", line 418, in clone
    agent=copy.deepcopy(self._tmpl_attrs.get("agent")),
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../copy.py", line 151, in deepcopy
    rv = reductor(4)
         ^^^^^^^^^^^
TypeError: cannot pickle 'TextIOWrapper' instances

Versions

• OS: macOS
• ADK version: latest (installed via pip install google-adk[mcp])
• Python version: 3.12.7

Additional context

• The MCP server is deployed on Cloud Run and is publicly accessible
• The same agent works perfectly when run locally with adk run or adk web
• This blocks production deployment of any ADK agents using MCP tools
• Attempted workarounds:
  • Using AdkApp wrapper (shown in code above) - still fails
  • Downgrading from Python 3.13 to 3.12.7 - still fails

Is there a recommended pattern for deploying agents with MCPToolset to Agent Engine? The current behavior seems to make it impossible to deploy any agent using MCP tools.