Add clustering evaluation with color toggle, ARI/NMI, and cluster analysis

[NetZissou]

Summary

Adds evaluation capabilities to the "Use column values" clustering mode, allowing users to assess how well KMeans clusters align with known categorical labels.

• Color toggle: radio in plot options to switch between KMeans cluster coloring and ground truth column coloring -- same points, same positions, instant comparison
• ARI/NMI metrics: Adjusted Rand Index and Normalized Mutual Information computed automatically after KMeans, displayed in the Cluster Analysis section
• Cluster analysis tree: full-width console output showing per-cluster composition with purity, entropy, and ranked breakdown by ground truth category
• Cardinality cap at 16: blocks clustering when column has >16 unique values to prevent misleading duplicate colors -- suggests filtering first
• Removed "Label by column" mode: redundant with the color toggle in "Use column values"
• Cleaned up unused run_dim_reduction/run_dim_reduction_safe from ClusteringService

Depends on #22

Generated with Claude Code

[egrace479]

It'd be great if we could color by other column values too -- maybe a dropdown?

Let's use a color scheme with 20 colors, and just have a warning of color duplication, instead of limiting based on the distinct color limitation.

[NetZissou]

Closing this PR in favor of a redesigned approach discussed in a follow-up meeting.

What's changing: The current design couples dim reduction and KMeans into a single "Run Clustering" button with mode switches. The redesign separates them into independent operations:

1. Dim reduction runs first (standalone) -- produces the 2D scatter plot
2. Color by any column via a general dropdown -- metadata columns are immediately available
3. KMeans is optional -- when run, it just adds another label to the color dropdown (since KMeans on full embeddings is just a mapping of observation -> cluster ID)

This better reflects the actual data flow (both operations independently consume the same full-size embeddings) and makes the UI more intuitive.

The evaluation features from this PR (ARI/NMI, cluster analysis tree, color comparison) will carry forward into the new implementation -- just triggered by the general color-by selector rather than a dedicated mode.

PR #22 (bug fixes) remains valid and should merge first.