12-10 Dual-Channel Magnitude-Phase Learning with Cross-Attention for Bone-Conducted Speech Restoration