Revert "Use virtual architecture to setup default CUDA_ARCH"

This reverts commit 8b7d0d05. An unintended consequence is that now only PTX is generated by default, unless CUDA_NVCC_FLAGS is used to generate code for specific devices.

@pcanal, what should we do in order to support PTX for a base arch, and binary for different arch? This might be useful when deploying to CVMFS with CUDA enabled.