[pytorch/pytorch] scaled_dot_product_attention_math usage and perf implications of returning the second value

### ROOT CAUSE The function `_scaled_dot_product_attention_math` returns two tensors (output and attention weights), but the main function `scaled_dot_product_attention` only uses the first one. This results in unnecessary computation and memory allocation for the second tensor (attention weights), which is not used by the main function. The performance impact is particularly noticeable in large-scale models where attention weights are not required. ### CODE FIX Change the function `_scaled_dot_product_attention_math` to return only the output tensor and remove the computation of the attention weights. This avoids the unnecessary computation and memory allocation for the unused tensor. ```cpp // Original function returning two tensors: // Tensor, Tensor _scaled_dot_product_attention_math(...) { // // ... compute output and attention_weights // return {output, attention_weights}; // } // Fixed function returning only one tensor: Tensor _scaled_dot_product_attention_math(...) { // ... compute output (without attention_weights) return output; } ``` This change ensures that the attention weights are not computed unless explicitly requested by other functions, improving performance.