[SYCL][E2E] Fix interop_all_backends for CUDA (#17643)

The CUDA backend has two different versions of the interop. A legacy
version and a newer experimental version. Using the experimental version
requires defining a macro before including the SYCL headers. This test
uses the experimental version so it should define the macro.

Fixes: #16070

Author: npmiller

sycl/test-e2e/Basic/interop/interop_all_backends.cpp

@@ -1,7 +1,5 @@
-// XFAIL: target-nvidia
-// XFAIL-TRACKER: https://github.com/intel/llvm/issues/16070
 // RUN: %if target-spir %{ %{build} -o %t-opencl.out %}
-// RUN: %if target-nvidia %{ %{build} -isystem %sycl_include -DBUILD_FOR_CUDA -o %t-cuda.out %}
+// RUN: %if target-nvidia %{ %{build} -DBUILD_FOR_CUDA -DSYCL_EXT_ONEAPI_BACKEND_CUDA_EXPERIMENTAL -o %t-cuda.out %}
 // RUN: %if target-amd %{ %{build} -DBUILD_FOR_HIP -o %t-hip.out %}
 
 #include <sycl/backend.hpp>