Default to UMAP dimensionality reduction method?

[PaulHax]

UMAP is much slower, but much more intesteing.

With 1000 images, dim reduction time in seconds:
PCA: 0.08
UMAP: 10.4

Perhaps we just need a faster/better way to "scrub" around the tightly clustered clouds made by PCA?

Ideas for copeing with slow embeddings+dim-reduction plot:

• Cache embedding computation.
• Allow pre-computing of UMAP, then Explorer loads it from "dataset."
• Spinner over whole plot, not just the "Compute" button.
• Instead of spinner, BS progress bar: image_count * seconds_per_image * 1/cpu_power