Create example multimodal training notebook

In the LSDB docs, we should create a nice example of training a multimodal dataset. It's fine if the API is cumbersome in some aspects, as we can use this notebook as a test case for a typical user flow which we can always make API improvements for down the road. (Thinking about things like needing to do repeated operations per cutout as opposed to a single operation for all cutouts, for example)

Review the LSDB docs and any existing notebook examples first, then trace the documented API for loading a multimodal dataset and operating on cutouts. Create a representative training workflow, including repeated per-cutout operations where needed, and confirm the notebook runs end to end and demonstrates a typical user flow.

Written by the indexing model from the issue text.