Fix division by zero in final avg improvement calculation

Fix category enum: use lowercase for ProbeAction categories

Fix eval_report.py: use step() return value and extract reward from observation

Add eval_report.py for before/after training comparison