Implement multi-head causal self-attention by stacking several independent single-head attention modules in parallel and concatenating their context outputs along the embedding dimension.
You are given an input tensor x of shape (B, T, d_in) and a list of H heads, where each head is a tuple (W_q, W_k, W_v) of weight matrices, each of shape (d_in, d_head). For every head independently:
x into queries, keys and values using that head's weights.(B, T, d_head).Finally, concatenate the H per-head context tensors along the last axis to produce an output of shape (B, T, H * d_head). Return a numpy array. Do not round inside the function.