Add more flexible FPU operator

[wardvermeulen]

This PR introduces more flexibility to the FPU operator as well as a ConfigError exception to handle misconfigurations of the kernel more easily.

The operator changed from GeneralFPUOp{M, N, K, CT, AT} to:

# mb, nb, kb are the base operator shapes (kb must be equal to 1 for now).
# M, N, K must be multiples of mb, nb, and kb respectively.
abstract type GeneralFPUOp{M, N, K, mb, nb, kb, CT, AT} end

Here, mb, nb, and kb are the internal base operator shapes that used to be equal to 4, 8, and 1 respectively by default.

[thomasfaingnaert]

CI failure is because I changed the benchmark keys to include the base operator shape.