Count ActionEvent tool calls in efficiency metrics

[DB825]

Summary

• Count SDK ActionEvent messages as tool calls in efficiency metrics using tool_name
• Preserve the existing assistant-message tool_calls path
• Clarify cumulative TokenEvent prompt-token accounting and update the metrics tests for the current flattened output

Why

CodeSearchGenerator serializes OpenHands SDK conversation events with event.model_dump() before calling compute_all_efficiency_metrics. Tool invocations in those serialized trajectories are represented as ActionEvent dictionaries, but compute_tool_call_metrics only counted assistant messages containing OpenAI-style tool_calls.

That could undercount tool usage for SDK event dumps and report avg_tool_calls_per_step as zero even when the agent used tools.

Tests

• python -m py_compile src\metrics\efficiency_metrics.py tests\metrics\test_efficiency_metrics.py
• python -m pytest tests\metrics\test_efficiency_metrics.py -q
• python -m pytest tests\metrics -q
• python -m ruff check src\metrics\efficiency_metrics.py tests\metrics\test_efficiency_metrics.py