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

### ROOT CAUSE The `scaled_dot_product_attention_math` function always returns both the output and attention weights, even when attention weights are not needed. This leads to unnecessary computation and memory overhead in scenarios where only the output is required. By introducing an optional parameter to control the return of attention weights, we can avoid these overheads. ### CODE FIX ```python def scaled_dot_product_attention_math( query: Tensor, key: Tensor, value: Tensor, scale: float = None, key_padding_mask: Optional[Tensor] = None, return_attention_weights: bool = False, ) -> Tuple[Tensor, Optional[Tensor]]: # Compute attention scores and weights scores = torch.matmul(query, key.transpose(-2, -1)) if scale is not None: scores = scores * scale if key_padding_mask is not None: scores = scores.masked_fill(key_padding_mask == 0, -10.0) attn_weights = torch.softmax(scores, -1) # Compute output attn_output = torch.matmul(attn_weights, value) if return_attention_weights: return attn_output, attn_weights else: return attn_output ``` This change adds a `return_attention_weights` parameter with a default value of `False`, allowing the function to return only the output when attention weights are not needed. This reduces memory usage and computation overhead in relevant cases.