Add periodic safe eval during training

[eugenevinitsky]

Summary

• Periodically run policy with safe/law-abiding reward conditioning (high penalties for collisions, offroad, overspeed) and render videos + collect metrics under the eval/ wandb tab
• Add [safe_eval] config section in drive.ini with configurable interval, episode count, and safe reward values
• Add --config flag to C visualizer so it can render with alternative INI files
• Refactor all 3 eval subprocess functions (wosac, human_replay, safe_eval) to share a common _run_eval_subprocess helper, eliminating ~50 lines of duplicated boilerplate

How it works

1. At safe_eval_interval epochs, copies the exported .bin model
2. Generates a temp INI with reward bounds at safe values
3. Renders videos via the C visualizer using the safe INI → logged under eval/world_state and eval/agent_view in wandb
4. Runs a Python subprocess (pufferlib.pufferl safe_eval) to collect metrics → logged under eval/* in wandb

Test plan

• Run training with safe_eval = True and verify videos appear in wandb eval tab
• Verify metrics are logged under eval/ prefix
• Confirm existing render and WOSAC/human_replay evals still work after the _run_eval_subprocess refactor
• Test with safe_eval = False (default for non-drive envs) to confirm no side effects

NOTE: code heavily generated by claude, this is just a prototype