[CUDA] Don't define __NVCC__.

[jlebar]

Summary

We decided this makes life too difficult for code authors. For example,
people may want to detect NVCC and disable variadic templates, which
NVCC does not support, but which we do.

Since people are going to have to change compiler flags *anyway* in
order to compile with clang, if they really want the old behavior, they
can pass -DNVCC.

Tested with tensorflow and thrust, no apparent problems.

cfe/trunk/lib/Headers/__clang_cuda_runtime_wrapper.h

 
 // Restore state of __CUDA_ARCH__ and __THROW we had on entry.
 #pragma pop_macro("__CUDA_ARCH__")
 #pragma pop_macro("__THROW")
 
 // Set up compiler macros expected to be seen during compilation.
 #undef __CUDABE__
 #define __CUDACC__
-#define __NVCC__
 
 #if defined(__CUDA_ARCH__)
 // We need to emit IR declaration for non-existing __nvvm_reflect() to
 // let backend know that it should be treated as const nothrow
 // function which is what NVVMReflect pass expects to see.
 extern "C" __device__ __attribute__((const)) int __nvvm_reflect(const void *);
 static __device__ __attribute__((used)) int __nvvm_reflect_anchor() {
   return __nvvm_reflect("NONE");