feat: add MCP tool call result callback

[seratch]

This pull request adds a tool_call_result_callback option for local MCP servers so applications can observe MCP CallToolResult metadata without changing the model-visible tool output. The callback receives an MCPToolCallResultContext with the original MCP tool name, public display name, parsed arguments, _meta, structuredContent, isError, and a copied tool output, which lets apps collect frontend-facing data for dual-output MCP tools while keeping replay and model history behavior unchanged.

This pull request resolves #3477.

from agents.mcp import MCPServerStreamableHttp, MCPToolCallResultContext


def tool_call_result_callback(context: MCPToolCallResultContext) -> None:
    app_context = context.run_context.context
    if not isinstance(app_context, dict) or context.result_meta is None:
        return

    app_context.setdefault("frontend_events", []).append(
        {
            "server": context.server_name,
            "tool": context.tool_display_name,
            "meta": dict(context.result_meta),
        }
    )


server = MCPServerStreamableHttp(
    name="Metadata-aware MCP",
    params={"url": "http://localhost:8000/mcp"},
    tool_call_result_callback=tool_call_result_callback,
)