Slim schema to 4 cols: voice_id, audio, text, source_clip_id

Fix audio schema to struct<bytes,path> + add text transcripts + HF feature metadata

Initial upload: 2000 voices from LibriTTS-R (longest clip per speaker, 12-30s)