Workaround nvfortran O3 bug in CMake Release build.

A workaround for #311 (closed) to pass more tests. Hopefully this change can be reverted once #311 (closed) is resolved.

before

The following tests FAILED:
	 40 - system--pw_atom--atom-pbe (Failed)
	 41 - system--pw_atom--atom-sigmapbe (Failed)
	 44 - system--pw_b3lyp--b3lyp-h2o (Failed)
	116 - system--pw_lattice-ibrav--lattice-wyckoff-sio2 (Failed)
	117 - system--pw_lda+U--lda+U+V-user_ns (Failed)
	120 - system--pw_lda+U--lda+U-noU (Failed)
	121 - system--pw_lda+U--lda+U-user_ns (Failed)
	122 - system--pw_lda+U--lda+U (Failed)
	123 - system--pw_lda+U--lda+U_background_one_channel (Failed)
	125 - system--pw_lda+U--lda+U_force (Failed)
	126 - system--pw_lda+U--lda+U_force_stress_ortho (Failed)
	128 - system--pw_lda+U--lda+U_kind1_collin (Failed)
	164 - system--pw_pawatom--paw-atom (Failed)
	166 - system--pw_pawatom--paw-atom_lda (Failed)
	167 - system--pw_pawatom--paw-atom_spin (Failed)
	168 - system--pw_pawatom--paw-atom_spin_lda (Failed)
	229 - system--pw_uspp--uspp1-coulomb (Failed)
	240 - system--pw_vdw--beef-spin (Failed)
	254 - system--pw_workflow_vc-relax_dos--vc-relax-dos-1 (Failed)
	255 - system--pw_workflow_vc-relax_dos--vc-relax-dos-2 (Failed)
	256 - system--pw_workflow_vc-relax_scf--vc-relax-scf-1 (Failed)
Errors while running CTest

after

The following tests FAILED:
        116 - system--pw_lattice-ibrav--lattice-wyckoff-sio2 (Failed)
        117 - system--pw_lda+U--lda+U+V-user_ns (Failed)
        120 - system--pw_lda+U--lda+U-noU (Failed)
        121 - system--pw_lda+U--lda+U-user_ns (Failed)
        122 - system--pw_lda+U--lda+U (Failed)
        123 - system--pw_lda+U--lda+U_background_one_channel (Failed)
        125 - system--pw_lda+U--lda+U_force (Failed)
        126 - system--pw_lda+U--lda+U_force_stress_ortho (Failed)
        128 - system--pw_lda+U--lda+U_kind1_collin (Failed)
        164 - system--pw_pawatom--paw-atom (Failed)
        166 - system--pw_pawatom--paw-atom_lda (Failed)
        167 - system--pw_pawatom--paw-atom_spin (Failed)
        168 - system--pw_pawatom--paw-atom_spin_lda (Failed)
        240 - system--pw_vdw--beef-spin (Failed)

There still quite many failures in CMake builds and most error are

     Error in routine c_bands (1):
     too many bands are not converged

I think this means wrong numbers in intermediate steps and this remains to be investigated separately from this MR.

This workaround is only needed in CUDA build, CPU build by nvfortran doesn't need to disable -O3.