V2 simplify

[himjl]

Summary

This PR simplifies the package around the actual product surface: benchmark data loaders, feature-driven simulation, and learning statistics.

It removes mref from the core learning-model and dataset-loading paths, flattens the linear learner implementation, and makes local data caching explicit and repo-local by default.

What Changed

• Flattened LinearLearner into hobj.learning_models and moved representation/update-rule modules to top-level hobj.learning_models.
• Replaced learner-facing image typing with local ImageId typing in hobj/types.py.
• Removed mref from learning-model code and tests.
• Refactored image dataset loading into hobj.data_loaders.images and removed the old hobj.data.images path.
• Reworked Imageset so it:
  • accepts cachedir: Path | None = None
  • defaults to repo-local data/
  • downloads/extracts assets directly without default_data_store
  • resolves local files by ImageId / Path instead of ImageRef
• Reworked hobj/data_loaders/behavior.py to cache downloaded JSON locally without default_data_store.
• Updated benchmark code to use the rewritten image/data loader interfaces.
• Switched CI to uv in .github/workflows/ci.yml and added uv.lock.
• Bumped supported Python to 3.12-only in pyproject.toml.

Why

The previous design exposed infrastructure abstractions like mref, ImageRef, and a global data store in places where the library really only needs:

• stable image IDs
• local cached benchmark data
• feature lookup for simulation

This PR narrows the API around those real needs and reduces coupling across the package.

Impact

• mref is no longer part of the core learning-model / data-loader workflow.
• Code importing old hobj.data... image modules will need to switch to hobj.data_loaders....
• Python 3.11 support is dropped; this branch targets Python 3.12.
• Local dataset caching now defaults to data/ in the repo rather than the previous global cache/store path.

Validation

• uv run pytest -q
• Current local result: 4 passed