Numerical stability improvements for Cosine Distance

[ptondereau]

Pull Request

Changes

• Added f32::EPSILON threshold check instead of strict zero comparison
• Added clamp() to ensure cosine stays within [-1, 1] range

Technical Details

The main improvement addresses potential floating-point precision issues:

• Previously, the code used pnqn != 0.0 which can be unreliable with floating-point arithmetic
• Now uses pnqn > f32::EPSILON to properly handle near-zero magnitudes
• Added protection against rounding errors that could push cosine outside [-1, 1]

PR checklist

Please check if your PR fulfills the following requirements:

• Does this PR fix an existing issue, or have you listed the changes applied in the PR description (and why they are needed)?
• Have you read the contributing guidelines?
• Have you made sure that the title is accurate and descriptive of the changes?

[irevoire]

Thanks for the PR, @ptondereau!

I ran the relevancy benchmarks:

Main

This PR

---

To me:

• The relevancy is pretty much the same; there are some differences here and there, but it’s probably just noise
• The indexing performances improved a little bit over all datasets
• Search performances and size on disk are pretty much the same as before

Overall, it’s good, so I’m merging 🦔