SNAP: Speaker Nulling for Artifact Projection in Speech Deepfake Detection