cleanup: strip utils.py to only do_mixup and interpolate

cleanup: remove dead visual/transformer classes, CLAPVisionCfg, dead CLAP methods from model.py; trim factory.py

cleanup: remove pann/timm/bert/bart dead branches from model.py

cleanup: trim open_clip to inference-only (remove loss, tokenizer, linear_probe, bert)

cleanup: remove dead code (training/, audio modality, chunk_inference, unused imports)

fix: remove unused h5py/tqdm imports from open_clip/utils.py

initial deployment of AudioSep to HARP