nx_vq_train.nx

nx_vq_train.nx -- Lloyd's algorithm (k-means) VQ codebook training: moonshot rung #2. Given a training set of LPC reflection vectors, it iteratively (1) assigns each vector to its nearest codebook entry and (2) moves each centroid to the mean of its assigned vectors -- provably non-increasing distortion. This is the optimal CLASSICAL codebook that neural RVQ generalises: it lowers the VQ distortion at the SAME bitrate (better coverage), the bridge from raw VQ to the learned codebook. Composes nx_vq. Caller owns all scratch (no hidden alloc). license_tier: ORIGINAL