Compiling with "non-default" CUDA

Hi,

when I specify a custom CUDA path when building, eg. `-DCUDNN_ROOT=/usr/local/cuda-8.0/`, the compilation toolchain still use parts which comes from the default CUDA version installed at `/usr/local/cuda/` (on my cluster).

Thus, when compiling, I got the following error:
`nvcc fatal : Value 'sm_61' is not defined for option 'gpu-architecture'`
because CUDA 7.5 does not support this architecture (which is the CUDA version installed at `/usr/local/cuda/`)

To fix this, I had to modify the PATH env. variables:
`export PATH=usr/local/cuda-8.0/:$PATH`

It would be nice if this workaround was not necessary.

Reproduce the build with CUDNN_ROOT set to /usr/local/cuda-8.0/ while another CUDA version is at /usr/local/cuda/. Inspect how the compilation toolchain selects CUDA paths and confirm that the requested version is used consistently, so sm_61 compiles without changing PATH.

Written by the indexing model from the issue text.