Add RunPod-ready Docker + startup (SSH/Jupyter/FileBrowser) with CUDA 12.1 + cuDNN 9

[Mars122]

Title: Add RunPod-ready Docker + Startup (SSH, JupyterLab, FileBrowser) with CUDA 12.1 + cuDNN 9

Summary

This PR adds an optional, self-contained Docker + startup setup for quickly running DiffSplat on GPU pods (e.g., RunPod). It includes:

• SSH server (22), JupyterLab (9999), FileBrowser (9090), optional Nginx welcome (80)
• CUDA 12.1 base with cuDNN 9 (libcudnn9-cuda-12)
• PyTorch 2.3.1 cu121 + xformers 0.0.27
• Clear, documented startup script with key-based SSH and Jupyter password support

Motivation

Many users run DiffSplat on cloud GPU pods. A batteries-included image with consistent ports and a one-shot startup script reduces setup friction, makes debugging easier (SSH/Jupyter/FileBrowser), and accelerates new contributors.

What’s included

• Dockerfile.runpod — Non-invasive Dockerfile that assumes it lives at repo root and installs common GPU + dev tooling, Jupyter, FileBrowser, SSH.
• start.runpod.sh — Starts the three services safely; emits useful debugging info.
• Docs (this PR body) — How to use and why.

Ports

• 22: SSH
• 80: Nginx (optional welcome page)
• 9090: FileBrowser
• 9999: JupyterLab

Auth

• SSH keys: set PUBLIC_KEY env var (or bake via build arg) to auto-append to ~/.ssh/authorized_keys.
• Jupyter: set JUPYTER_PASSWORD (recommended) or copy token from logs.

RunPod notes

• HTTP services (80/9090/9999) appear under https://<pod>-<port>.proxy.runpod.net/.
• SSH is a raw TCP port; map it in the UI and connect with ssh -i <key> root@host -p <port>.

Testing steps

1. Build the image and run with GPU access.
2. Confirm nvidia-smi works in the container.
3. Open Jupyter and FileBrowser via the proxy URLs.
4. Use SSH + SFTP to move files quickly.

Non-goals

• This does not change core DiffSplat code, training loops, or dependencies beyond runtime tooling.
• The Dockerfile is optional; existing workflows continue to work.

License and attribution

• Apache 2.0 (matches project license).
• Adapted from the author’s GPU dev image (Mars122/DiffSplat-Docker).