Implement Multi-Head Latent Attention (MLA)

Implement Multi-Head Latent Attention (MLA), and return both the projected attention output and the compressed latent tensor.