Harden evals with retry logic + centralize logs on Docker host

[cte]

Description

Add retry logic for unexpected container failures (due to CPU or RAM spikes).

---

Important

Adds retry logic for container failures and centralizes logging to Docker host using a new FileLogger class.

• Behavior:
  • Adds retry logic in processTaskInContainer() in processTask.ts for container failures, with exponential backoff up to 10 retries.
  • Centralizes logs to /var/log/evals on Docker host by modifying docker-compose.yml and runs.ts.
• Logging:
  • Introduces FileLogger class in FileLogger.ts for logging with levels (INFO, ERROR, WARN, DEBUG).
  • Updates runEvals(), processTask(), and runTask() to use FileLogger for logging.
• Misc:
  • Modifies createRun() in runs.ts to stop removing Docker containers automatically.
  • Updates docker-compose.yml to mount /tmp/evals to /var/log/evals for logging.

^{This description was created by for a3c55bb. You can customize this summary. It will automatically update as commits are pushed.}

[ellipsis-dev]

To ensure FileLogger cleanup even on non-timeout errors, consider moving logger.close() into a finally block so it always executes.