Reshape

Pick the formula that defines head_dim from hidden_size and num_heads.

Walk through how a token's d_model embedding is split into per-head pieces inside multi-head attention.