VLM: Wire into preprocessing pipeline as a vlm_step

[madhavcodez]

The preprocessing pipeline currently runs parse_step → load_step. The VLM classification should slot in between as parse_step → vlm_step → load_step.

What this looks like:

1. Create util/preprocessing/vlm_step.py with run_vlm_step(parsed_json_path)
2. Read parsed_data.json, iterate image pairs + locations
3. Call Gemini for each pair, write the result into the "prediction" field (already null in the JSON — parse_step.py line 187 sets this up)
4. Save the updated JSON
5. Add "vlm" to the order list in preprocess-data.py line 56 so --start-at vlm / --stop-after vlm works

The chat.vlm_assessments table from readme_Chat.md is the eventual DB destination but getting predictions into the JSON first is the right step — easier to inspect before committing to the database.

546 image pairs and 11,548 locations are loaded and ready to test against.