fix(tools): prevent duplicate keyword arg when MCP tool param name shadows injection key

[s-zx]

Summary

Fixes #6760.

FunctionCall._build_entrypoint_args() unconditionally injects the framework Agent and Team objects into entrypoint_args for any entrypoint whose signature contains parameters literally named agent or team. The MCP call_tool wrapper function intentionally uses these names as optional framework-injection parameters via **kwargs forwarding. However, when the actual MCP tool being called (e.g. Linear's save_issue) also exposes a parameter named team (as a normal string/ID parameter), the LLM provides it in self.arguments. The subsequent entrypoint(**entrypoint_args, **self.arguments) call then has team in both dicts, raising:

TypeError: functools.partial(...) got multiple values for keyword argument 'team'

Fix: Before spreading both entrypoint_args and self.arguments into the entrypoint call, remove from entrypoint_args any key that already exists in self.arguments. This lets the LLM-provided value win and eliminates the collision. The change applies to both the sync (execute) and async (aexecute) code paths in FunctionCall.

No regression risk: For normal (non-MCP) tools, self.arguments never contains framework-reserved names (agent, team, run_context, etc.) because the LLM only sees user-defined parameters. The deduplication only takes effect when there is an actual name collision.

Test Plan

• Create an agent with MCPTools connected to a server that exposes a tool with a team parameter (e.g. Linear's save_issue)
• Verify the tool call succeeds without TypeError: got multiple values for keyword argument
• Verify that regular agents with tools using agent: Agent / team: Team injection still work correctly