add FP64 support for attention kernels

[azrael417]

It would be good if our attention layer would support fp64 input. This needs to be reviewed carefully, as the current kernels hard-code fp32 in various places.