fix: show tool names in history and return the model answer

[anxkhn]

Fixes #438.

Agent.stream stored each MCP call as {'tool': name, ...}, but called_tools_history.jinja reads tool.name. The next decide prompt showed a blank tool name, so the agent could not see what it already called.

When no further tools were needed, the complete event also replaced the last model response with Task completed. Clients that read the final artifact got the placeholder instead of the <Answer> body.

This change stores the history record under name (and includes isError in the template), keeps the last model response as the complete content, and stops emitting Step N / raw history strings as artifact text. Tests render the real Jinja template and check the complete event.

Tested with:

uv run --with pytest --directory samples/python/agents/a2a-mcp-without-framework pytest tests/test_tool_history.py -q

[soyeladice-svg]

Burs-IA technical review (operator-assisted; public-source only):

I reproduced one boundary condition in the current PR logic: at the ten-call limit, the loop can reach its success path while the model response still contains a tool/function request rather than a terminal answer.

A focused regression would be useful here: make the 10th model response request another tool and assert that the run does not report success/terminal completion while an unresolved tool request remains.

This is a bounded review finding, not a conformance claim. No private endpoints or credentials were used.