add: Choosing fusion method notebook

[Dylancouzon]

Collab link: https://colab.research.google.com/github/qdrant/examples/blob/fusion-methods-tutorial/fusion-methods/Choosing_a_Fusion_Method.ipynb

Summary

Adds fusion-methods/Choosing_a_Fusion_Method.ipynb, the runnable companion to the hybrid-queries
docs in qdrant/landing_page. The notebook walks through choosing between Qdrant's fusion methods
(RRF, weighted RRF, DBSF, FormulaQuery) with real metrics on a real corpus.

Running the notebook

The notebook is designed for Google Colab. It reads QDRANT_URL and QDRANT_API_KEY from Colab secrets via from google.colab import userdata. To run it
locally instead, swap the credentials block in the imports cell for os.environ or getpass.

Decisions worth noting

• BEIR/SciFact as the eval corpus over larger BEIR datasets (NFCorpus, FiQA): fits a Free Tier cluster and finishes in a reasonable runtime, with enough relevance labels (300 queries, ~339 query-doc pairs) for the metrics to be meaningful.
• Grid search for weighted-RRF weights, not random search or Bayesian optimization. For a 2D weight space with a smooth metric surface, all three land within rounding error; grid is the simplest to explain and reproduce.
• No train/val split in the notebook itself for simplicity. The docs page and the notebook's "Best practice" callout direct readers to apply a held-out split for production tuning.
• 0.1 coefficient on the decay term in the FormulaQuery demo. RRF scores in the top-5 are roughly 0.2-0.5 while un-weighted decay returns [0, 1].
  Without the coefficient, recency crowds out relevance and the demo silently replaces the fused candidates instead of nudging the ranking.