Add int8 (8-bit k-means-palettized) weight-only tier of both encoders

clapkit T1: LAION CLAP both encoders, fp16 (spectrogram-input audio + RoBERTa text), pre-norm

initial commit