Add causal person fusion and MuVAP stage-3 baselines

requirements.txt

@@ -1,4 +1,5 @@
 torch
+torchaudio
 numpy
 # Default for Thinker replay. The B1 local extractor also supports 4.57.6;
 # keep that version in existing GPU containers (see docs/local_feature_baseline.md).
@@ -9,6 +10,8 @@ librosa
 Pillow
 tqdm
 scikit-learn
+PyYAML
+einops
 
 # Used by train_omni_turn_lora.py; already included in the ms-swift image.
 peft