[CMake] fix GGA-noncolin in GPU builds due to miscompliation

The CMake link to openacc in FFTXlib was missing (my fault). This caused failure in CMake runs with GGA-noncolin, where the fftx_threed2oned_gpu routine was called from FFTXlib.
However, by fixing this, my cmake-nvhpc pipeline fails when it comes to compile the c files in FFTXlib with the following message:
nvc-Error-CUDA 9.0 or later required
To fix this (thanks @pietrodelugas ) you need to impose the cuda version 11.2 in the .yml file (2nd commit). However this is just a provisional fix. Is it a problem related to nvhpc 21.2 only?
@ye-luo