Teaching Audio Models to Reason: A Unified Framework for Source- and Layer-wise Distillation