[CUDA] Collapsed offload actions should not be top-level jobs.

If they are, we end up with the last intermediary output preserved
in the current directory after compilation.

Added a test case to verify that we're using appropriate filenames
for outputs of different phases.

Differential Revision: https://reviews.llvm.org/D23526

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@279455 91177308-0d34-0410-b5e6-96231b3b80d8