Add timings for phases

[jeffsee55]

Starting experiment...
Starting 1 eval attempts concurrently (1 evals × 1 runs)
Running smoke/opus/_smoke [1/1] [vercel-ai-gateway/claude-code · opus]...
  ... Creating sandbox (0.5s)
  ... Uploading files (2.2s)
  ... Initializing git (0.5s)
  ... Running setup (26.5s)
  ... Installing dependencies (14.5s)
  ... Running agent (183.0s)