Explicit make targets for cuda, cppnone, cppsse4 etc (avoid building CUDA code 5 times for each C++ AVX mode)

[valassi]

Part of the cleanup of builds. Avoid building CUDA code 5 times for each C++ AVX mode. This is a bit the mirror of #679: CUDA builds should not depend on C++ builds and viceversa

[valassi]

Note: in MR #723 fixing #725, I improved the separation of cpu and gpu namespaces (so now it is a bit sfare to mix the two codes... though it is maybe a better idea not to do that anyway). Maybe eventually we need separate namespaces for all SIMD modes (if we go to fat binaries...).

[valassi]

I have added a long todo list in #602. When that task is done, it will also solve this #680.

[valassi]

Like #602, this will be fixed by PR #798.