* Implement scaled_dot_product_attention and test * Support attn_mask * Support is_causal too * Use in llama * Don't forget to reshape * Set requires_grad=False for causal * Remove staticmethod * Remove extra spaces