Handle empty EVMbench summaries

[sylvesterkaczmarek]

Summary

• make EVMbench.get_full_summary() handle an empty result set
• use the eval's own run_group_id rather than indexing the first result task
• preserve existing score, health, and per-audit aggregation behavior

Nanoeval's base Eval.self_test() explicitly calls get_full_summary([]). EVMbench currently builds a task list and then reads tasks[0] solely to populate the run-group ID, which raises IndexError when no results are available.

The eval already owns the canonical run_group_id, so the summary can use that value directly without depending on a completed task.

Regression coverage verifies an empty summary reports zero samples and scores while retaining the eval run-group ID.