Evals for harnesses running in sandboxes.

Easily run evals across different sandboxes, harnesses and models to optimize performance, latency and costs.