[CUDA] Don't pass top-level -march down to device cc1 or ptxas.

Summary:
Previously if you did e.g.

  $ clang -march=haswell -x cuda foo.cu

we would pass "-march=haswell -march=sm_20" down to the ptxas tool.
This causes it to assert, and rightly so!

Reviewers: tra

Subscribers: cfe-commits, echristo

Differential Revision: http://reviews.llvm.org/D21419

llvm-svn: 272857