Normalize per-sample embeddings before averaging centroid

[ComputelessComputer]

Summary

Speaker embeddings must be L2-normalized before averaging so samples with larger raw magnitudes (typically longer or louder clips) don't bias the centroid. The previous normalizedEmbeddingCentroid summed raw WeSpeaker outputs and only L2-normalized the result at the end.

What changed

src-tauri/swift-permissions/src/speech_bridge.swift — per-sample L2 normalization before summation inside normalizedEmbeddingCentroid. Zero-magnitude samples are skipped. The final L2-normalization of the summed vector is preserved.

Why it helps

Centroid embeddings drive speaker similarity comparisons (used today for cross-meeting speaker identification and, after #5, for constraining over-segmented diarization). Giving each sample equal weight — regardless of raw magnitude — matches the standard recipe for averaging speaker embeddings and reduces drift when a speaker has one long monologue plus several short contributions.

What's not in this PR

• constrainDiarizedSegments embedding-based reassignment (separate PR).
• Stratified sampling across segments in selectSpeakerEmbeddingSegments (H3 in the issue) — follow-up.

Testing notes

Swift-only change. bun run build for the frontend still passes. Please verify with the existing Swift test suite and, if available, the in-app speaker-suggestion flow with a known speaker to confirm match quality is the same or better.

Addresses #4.