cp.x's Conjugate Gradient routine on GPU + little clean up of the code

it looks like it may work

It works in the norm conserving case. Currently using it with a big system. Some performance optimization is still possible

before merging: I have to clean up the commits and maybe remove some debug facilities

some useful debug facilities are left there (they do nothing if not asked to) for future use (e.g. porting of ultrasoft/ ensemble dft to gpu)

TODO: test ensemble dft, probably it is broken

when ultrasoft or ensemble dft are used in the CG routine with GPU, an error is raised and the code terminates