Inaccurate rubric timings when using RubricGroup

[pjo256]

RubricGroup.score_rollout and score_group maintain reward and metrics between rubrics, but not timing. After N rubrics:

• scoring_ms = the last rubric's scoring time
• each rubric sees the cumulative total_ms, including other rubric's scoring_ms
• generation_ms + scoring_ms ≠ total_ms for a rubric group

This affects the following environments: MultiTurnEnv, ToolEnv, PythonEnv, SandboxEnv.