voice: fold to one canonical noise-sample generator (single source-of-truth)

[drewdrewthis]

Today two byte-deterministic noise-sample generators ship: TS (javascript/scripts/generate-noise-samples.mjs) and Python (python/scripts/generate_noise_samples.py).

After #587 (Python parity) lands and Python gets the richer generator too, fold to ONE canonical generator. Likely Python (keeps the rest of the asset-gen tooling co-located in python/scripts/). The JS workflow calls it via Make or uv run.

Two generators for the same deterministic byte-output is a guaranteed future-divergence.