[clang] cuda clang: Move nvptx-surface.cu test to CodeGenCUDA (PR #134758)

[Artem Belevich via cfe-commits]

================
@@ -2,6 +2,170 @@
 // RUN: %clang_cc1 -triple nvptx64-unknown-unknown -fcuda-is-device -O3 -o - %s -emit-llvm | FileCheck %s
 #include "Inputs/cuda.h"
 
+struct char1 {
----------------
Artem-B wrote:

Those are actually *useful* failures and expose real issues in those tests.

- correctly-rounded-div.cu defines their own float4. Should not do that for CUDA tests. Rename locally.
- offloading-entries.cu happens to define part of surface-related  data structures that your change is defining. That should also be reconciled. It may just work if you remove local declarations and adjust checks appropriately.
- profile-coverage-mapping.cu - looks cosmetic. We've added more functions, so there are more coverage entries.
- propagate-attributes.cu Ouch. Including a C++ header under "extern C" is not the best idea. The test only cares about the names of the functions in the test, so it's just a matter of making sure that `extern "C"` applies only to them.

I think it's a good opportunity to clean things up. If I didn't miss the mis-placement of the test during the initial review, it would all pop up during the review. We're just doing it a bit later. 

https://github.com/llvm/llvm-project/pull/134758