Break possible deadlock on lockstep execution

[fel-cab]

Some of the target_requires_atomic tests get on infinite loop with LLVM/Clang compiler because lockstep execution on the target device and the if-else order of execution. These tests are written assuming that target threads are independent or that if they execute in a warp, in lockstep execution, the order of execution would be first the if part and then the else part.
The modifications on the tests are meant to make sure that this problem does not happens for LLVM/Clang or any compiler that could encounter this problem.

[fel-cab]

The following Issue was created for LLVM issues llvm/llvm-project#94047
Given the conclusion by the developers. I think this PR should be closed, but the problem with LLVM will always remain. The original code doesn't pass on LLVM either.

This code cannot reliably work. You are trying to synchronize within a warp on an AMDGPU, that does not work. OpenMP 6.0 makes this explicit. If you want to test atomics, split the reader and writer into different warps at least.