docs: interpret custom augmentation embeddings

[iamdhrv]

Change

• explain how to compare the example labels with their nearest-neighbor label
  distributions
• interpret all three rendered neighborhoods
• distinguish the strong normal-image grouping from the weak multi-label
  abnormal match
• state the limitation of drawing an accuracy conclusion from three examples

Why

The tutorial says the plots evaluate representation quality but previously
stopped immediately after rendering them. The added interpretation answers
whether the displayed embeddings are good or bad, points readers to the
evidence in each plot, and frames this as a qualitative check rather than a
diagnostic metric.

Checks

• uvx ruff check docs/source/tutorials_source/package/tutorial_custom_augmentations.py
• git diff --check

Fixes #581