Don't use __restrict__ for CUDA

[StephanTLavavej]

#5079 was a nice try, but it doesn't work. We need to revert this before 17.13 Preview 3 branches for release.

It broke NVIDIA's Cutlass again. @CaseyCarter observed:

In both the old and new preprocessed repros, our _RESTRICTs in the declaration of _Adjacent_difference_no_overlap are being expanded to plain restrict. If I had to guess, I'd say that CUDA is replacing __restrict__ with restrict under the expectation that __restrict__ is only used in C mode.

Fixes VSO-2295101 / AB#2295101 again.

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.