feat: Improve TF-IDF weighting, match explanations, and MovieLens enrichment

[LavishSphere]

Summary

• Replaced weighted tfidf_soup repetition with separate TF-IDF indexes for description, genres, cast, and director
• Added explicit TF-IDF field weights:
  • Description: 50%
  • Genres: 30%
  • Cast: 10%
  • Director: 10%
• Integrated MovieLens 32M metadata into the data loader
  • Adds MovieLens genres
  • Adds user tag signals
  • Adds cached aggregate rating signals via ratings_summary.csv
• Updated TF-IDF input data so MovieLens genres and tags contribute to recommendation matching
• Updated recommender scoring so TF-IDF content similarity has more influence than IMDb/popularity
• Added match_breakdown to recommendation results so the backend can explain how match scores are calculated
• Disabled MDP diversity re-ranking for now so final results stay ordered by relevance score
• Updated README data notes to reflect included MovieLens assets