FuseMax: Leveraging Extended Einsums to Optimize Attention Accelerator Design