feat(python): expose agent run trace

[RitwijParmar]

Why

This addresses part of the long-running request in #262 around seeing which MCP tools were actually called, how many times, and what each tool returned before the model turns those results into a final answer.

Right now users can watch stream updates, but after a run there is no simple object they can hand to a dashboard, debugger, or notebook. That makes it harder to answer questions like: did the model call the same tool twice, did the tool return the expected raw payload, or did the final answer rewrite a tool result in a misleading way?

What changed

• adds AgentRunTrace and ToolCallRecord for the Python agent
• records tool name, call id, input, raw output, bounded output preview, timing, counts, final output, and errors
• exposes agent.get_last_run_trace(), agent.get_last_tool_calls(), and agent.get_last_tool_outputs()
• closes unmatched tool calls with a clear trace error instead of leaving them pending
• documents how to inspect tool calls after agent.run(...) or agent.stream(...)
• adds a unit test covering counts, inputs, outputs, and helper methods

I kept this as an inspectable in-memory trace instead of changing the streaming contract, so existing callers should not see behavior changes.

Checks

• uv run --extra dev pytest tests/unit/test_agent.py -q
• uv run --extra dev ruff check mcp_use/agents/__init__.py mcp_use/agents/run_trace.py mcp_use/agents/mcpagent.py tests/unit/test_agent.py
• uv run --extra dev ruff format --check mcp_use/agents/__init__.py mcp_use/agents/run_trace.py mcp_use/agents/mcpagent.py tests/unit/test_agent.py
• uv run --extra dev python -m py_compile mcp_use/agents/__init__.py mcp_use/agents/run_trace.py mcp_use/agents/mcpagent.py tests/unit/test_agent.py