The ~88 fine descriptors were re-encoded through RoBERTa on every folder's
forward — a fixed cost that made batch size irrelevant. Now cache the normalized
text embeddings + logit scale at load; per-folder forwards run only the audio
tower (get_audio_features) and a matmul against the cached text embeds. ~1.8x
(14s->7.7s/folder; text was ~45% of per-folder cost).
API note (transformers 5.10.2): get_text/audio_features return a model-output
object whose .pooler_output IS the projected 512-d joint embedding — verified
identical to a full ClapModel forward to 1.6e-7. No classification change.
