Add experiment progress and output logging

[Copilot]

Experiment runs now report queue progress, capture command output as artifacts, and avoid writing run output directly to stdout unless requested. Error-level run logs still surface to aid debugging.

• Progress events

  • Report completed, running, remaining, and total treatment counts while experiments run.

• Output hooks

  • Add sandbox command output hooks so callers can observe stdout/stderr without implicit process writes.
  • Route treatment logs through runner events with info, error, and raw output event types.

• Log artifacts

  • Persist per-treatment command output to:
    • stdout.log
    • stderr.log

• CLI output control

  • Add --show-output to mirror command output to stdout.
  • Prefix mirrored output with experiment, treatment, eval, model, and stream context.
  • Always emit error logs regardless of --show-output.

agent-eval --experiment mcp --concurrency 2 --show-output
# [MCP | Control | 001-agent-uses-button-from-primer | gpt-5.5 | stdout] ...