Investigate if we can remove `metadata` from datasets

Nobody has claimed this yet.

Given that we end up with a HyraxQL (GraphQL-like) method for defining particular aspects of a dataset to utilize during training, inference, and visualization, could we further simplify the dataset classes by treating the metadata tables in a similar way?

The HyraxQL expectation for a dataset is that it will expose a set of getters, such as `get_image`, `get_label`, etc. Currently there is also a `self._metadata_table` that can be used to pull information from. IIRC this was originally implemented for visualization in order to display tabular data when selecting points interactively.

One can imagine a scenario where a metadata table containing say photo-z estimates would be useful for training as well as visualization. So we _could_ automatically generate getters for all the columns that are found in the metadata table, and make those available for use in `to_tensor`.

Start by tracing the dataset classes, the _metadata_table access, and the HyraxQL getters such as get_image and get_label through to_tensor and visualization use. Determine whether metadata-table columns can be exposed as getters without breaking training, inference, or interactive visualization. Done means a documented design decision on whether metadata can be removed and what replacement behavior is required.

Written by the indexing model from the issue text.