Compute the attention output of shape (T_q, num_query_heads, head_dim). Each query head h_q reads from kv head h_q / (num_query_heads / num_kv_heads).

Row-major strides for a list of dim sizes.

Chars that appear on both sides of the equation but not in the output are summed away.

Computes the contraction value for one fixed output position. The output position is already encoded in `pos` (mutable scratch space shared across cal

Write an std::string to std::cout.