Add Colab quickstart notebook

[CodeSarthak]

Adds a self-contained colab_quickstart.ipynb at the repo root that runs a pretrained NanoWM checkpoint on free-tier Google Colab end-to-end. Linked from the README's Quick Start section.

Motivation

Getting from git clone to a working rollout on a fresh machine currently hits several rough edges that aren't covered by environment.yml alone — modern huggingface_hub / transformers versions break the pinned diffusers==0.24.0 imports, the released model.safetensors doesn't load via the pickle-based find_model, PyTorch >= 2.6 changed the torch.load(weights_only=...) default, etc. The notebook bakes all of these into a single top-to-bottom flow targeted at free-tier Colab (T4 GPU, 15 minutes total).

What it does

1. GPU sanity check
2. git clone the repo
3. pip install a pinned dep set (huggingface_hub==0.20.3, transformers==4.36.0, diffusers==0.24.0, accelerate==0.25.0, plus the rest from environment.yml)
4. One required runtime restart (Colab pre-imports huggingface_hub at session start)
5. Idempotent patch for find_model to pass weights_only=False (see find_model fails on PyTorch >= 2.6 (weights_only default change) #10 / linked PR for the upstream fix; the cell becomes a no-op once that lands)
6. snapshot_download the Point Maze B/2 checkpoint
7. Convert model.safetensors → model.pt so find_model can load it
8. Unzip the OSF Point Maze dataset (instructions for the manual download up top)
9. Run src/sample/rollout.py with known-good DINO-WM Point Maze flags (--history_length 3, --rollout_length 20, --num_sampling_steps 15)
10. Display the resulting compare.mp4 inline

Each cell is idempotent so a user can re-run individual steps after a transient failure without resetting the session. A troubleshooting table at the bottom maps each common error to which step to re-run.

Tested

Ran the notebook top to bottom on free-tier Colab T4 — produces the expected gen / gt / compare mp4s in ~15 minutes total wall-clock.

Notes

• Targets Point Maze (smallest demo). Swapping the HF repo id in step 6 + the OSF folder in step 8 generalizes to PushT / Wall / Rope / Granular.
• CSGO is intentionally out of scope (~675 GB dataset, not feasible on free Colab).
• The notebook references the find_model weights-only patch issue (find_model fails on PyTorch >= 2.6 (weights_only default change) #10). Once that PR lands, the patch cell will no-op cleanly — no churn needed.