Add MineRL Treechop transition dataset builder

[mkdsrt10]

Summary

Adds MineRLTreechop, a transition-level builder for locally acquired MineRLTreechop-v0 human demonstrations. This follows the redirect from galilai-group/stable-worldmodel#313 so the MineRL integration lives with the dataset library.

• caches one transition per row: RGB POV image, deterministic flattened action, reward, terminal flag, episode ID, trajectory name, and timestep
• retains temporal boundaries explicitly so world-model code can create its own windows
• imports MineRL lazily; it does not download or redistribute MineRL demonstrations
• includes a dedicated documentation page and mock-pipeline tests, requiring no Minecraft runtime in CI

Validation

• .venv/bin/python -m pytest -q stable_datasets/tests/video/test_minerl.py stable_datasets/tests/video/test_moving_mnist.py stable_datasets/tests/video/test_something_something_v2.py (9 passed)
• .venv/bin/python -m ruff check stable_datasets/video/minerl.py stable_datasets/tests/video/test_minerl.py

I intentionally kept this focused on the Treechop task and the dataset contract. Other MineRL/BASALT tasks can reuse the same pattern in follow-up contributions.

[mkdsrt10]

Follow-up validation on real locally acquired MineRLTreechop-v0 data: the builder discovered 209 trajectories and completed an end-to-end cache build for one actual trajectory (2,122 transition rows). The cached rows had 64x64 RGB observations and the expected 10-dimensional flattened action vector. I also added explicit cleanup for MineRL’s legacy worker pool after cache construction.