The CUDA Toolkit's VS integration provides abstractions for host
compiler options for `nvcc` to pass through `-Xcompiler` to the host
MSVC.  Populate our secondary flag table and use it to remove flags from
the `AdditionalCompilerOptions` in favor of their abstractions.

Unfortunately a bug in the CUDA 8.0 VS integration prevents us from
passing anything in `AdditionalCompilerOptions` reliably.  After taking
out the flags that have dedicated abstractions, drop the rest.