Publish reusable Modal runtime image - #1122

Summary

Make the remote-VM runtime reproducible and reusable through a prebuilt Docker image published manually to Docker Hub.

Why

Building the full CUDA, PyTorch, Open3D, and GTSFM dependency stack during every Modal deployment is slow and error-prone. A prebuilt runtime reduces repeated downloads and makes deployment behavior easier to inspect.

Changes

• Add a maintained Modal CUDA runtime Dockerfile
• Add the manual Docker Hub publishing helper
• Preserve CUDA package indexes and system compression dependencies
• Install dependencies in the runtime virtual environment
• Keep future Modal Volume mount paths empty during image extension
• Document the image publication workflow

Validation

• Python compilation of visualization/modal_app.py and visualization/modal_deployment.py
• git diff --check

Stack

Depends on #1121. This draft targets studio/package-cli so the review contains only the runtime-image changes. It should be retargeted to master after #1121 merges.