Efficiently Dispatching Flash Attention For Partially Filled Attention Masks