[2406.06567] DHA: Learning Decoupled-Head Attention from Transformer Checkpoints via Adaptive Heads Fusion