Fix inverted information diversity score (cosine distance vs. similarity)

[sundy1994]

Overview

Corrects the information diversity calculation in calculate_ID_score(), which was computing the opposite of topical diversity.

Description

The score implements Eq. (1) of Riedl & Woolley (2017): ID = Σ(1 − cos(dⱼ, M))² / N, where cos is cosine similarity. The code used (1 - cosine(doc, mean))**2, but scipy.spatial.distance.cosine() already returns the cosine distance (1 − similarity) — so the 1 - canceled it back to similarity. The result was the mean squared similarity to the mean topic vector, which moves opposite to diversity (high score = less diverse), contradicting the docs.

Changes

1. Drop the spurious "1 -".
2. Update docstring to describe squared cosine distance and note the scipy sign convention.
3. Update the dead source link → https://ssrn.com/abstract=2384068.
4. Remove 7 unused imports.

[xehu]

Simple fix. Looks good to me!