pub fn split_heads( qkv: &Tensor, n_head: usize, ) -> Result<(Tensor, Tensor, Tensor)>
qkv [t, 3c] -> (q, k, v) each [h, t, c/h].