Improve packaging

[himjl]

Summary

This PR repackages the hobj datasets into a simpler packaged-data layout and updates the library to consume that format directly. The main change is a move away from bespoke image-set/session objects toward pandas-backed loaders for behavior and image metadata.

What Changed

• Added packaged CSV-based behavior loaders for the high-variance and one-shot experiments.
• Added packaged image manifest loaders for highvar, oneshot, warmup, and catch datasets.
• Added direct image loading by image_id via load_image(...).
• Added a dataset download utility and CLI entry point: hobj-download-data.
• Updated benchmarks to build from the new packaged behavior/image metadata instead of the old loader classes.
• Updated hobj top-level exports to expose the new loading API.
• Updated tests and example notebooks to reflect the new packaged-data workflow.
• Removed older image loader modules and cache-oriented config that are no longer needed under the packaged layout.

API Direction

• Behavior loaders now return trial-level pandas.DataFrames.
• Image manifest loaders return metadata only.
• Actual image access is done through load_image(image_id).

Why

• The new format is easier to inspect, validate, and reuse outside the benchmark internals.
• It reduces custom data-container code and makes the package more interoperable with standard Python data tooling.
• It creates a clearer separation between metadata access and image loading.