In NLP, sequences in a batch are often padded to a common length using a special pad token id. During attention computation, padding positions should be ignored. Implement a function build_attention_mask(token_ids, pad_token_id) that takes a 2D NumPy array of token ids of shape (batch_size, seq_len) and an integer pad_token_id, and returns a 2D NumPy integer array of the same shape where positions containing real tokens are marked with 1 and padding positions are marked with 0. The returned array must have dtype int and be converted to a nested Python list via .tolist() for display.