Implement Multi-Head Attention (forward pass)

Task Implement the forward pass of Multi-Head Self-Attention (forward only; no backprop). Given: number of heads h sequence length n model dimension d_model per-head dimension d_head = d_model / h (guaranteed divisible) input X of shape (n, d_model) weight matrices W_Q, W_K, W_V, W_O, each (d_model, d_model) Compute: Q = X @ W_Q, K = X @ W_K, V = X @ W_V Reshape Q,K,V to (h, n, d_head) For each head: scores = (Q @ K^T) / sqrt(d_head) with shape (n, n) apply row-wise softmax to get attention weights A head_out = A @ V with shape (n, d_head) Concatenate all heads to (n, d_model) and multiply by W_O to produce Y of shape (n, d_model) Input (stdin) Line 1: n d_model h Next n lines: matrix X Next d_model lines: W_Q Next d_model lines: W_K Next d_model lines: W_V Next d_model lines: W_O Floats are space-separated. Output (stdout) Print Y as n lines of d_model floats, each with 6 decimals. Constraints 1 <= n <= 20 1 <= d_model <= 32 h divides d_model Use numerically stable softmax (subtract row max). Example Input 1 2 1 1 2 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 Output 1.000000 2.000000