perf: parallelize VLM image preprocessing across threads

[cdreetz]

Summary

• Parallelize VLM image preprocessing chunks using ThreadPoolExecutor (up to 8 workers) instead of processing sequentially
• PIL/numpy operations release the GIL, so threads give real concurrency for the CPU-bound image processor work
• Add unique image count to VLM timing log for better observability

Context

With CUA browser environments (64 rollouts, 8 per example, 15 turns with screenshots), the orchestrator processes hundreds of unique images through the Qwen3-VL processor. At chunk_size=32, this created 30+ sequential chunks taking ~17s total. Threading should cut this significantly on multi-core machines.

Test plan

• All existing trajectory tests pass (41/41)
• Verify reduced preprocessing time in VLM training runs

🤖 Generated with Claude Code

---

Note

^{Cursor Bugbot is generating a summary for commit c8c3406. Configure here.}