[SYCL][E2E] Stop CMAKE_CXX_FLAGS cache variable passing to clang++

[Maetveis]

Before #9075 the value of CMAKE_CXX_FLAGS was passed to SYCL E2E compilations. In #9075 the variable was explicitly unset in non-standalone (in-tree) builds of the tests. Unsetting a normal variable however exposes the cmake cache variable of the same name.

I believe this was not the intent, rather the intent was to not pass any CMAKE_CXX_FLAGS to the E2E tests in non-standalone builds. It is weird to pass the user's CMAKE_CXX_FLAGS to the E2E tests, as these flags are for the SYCL toolchain, not tests.

Passing CMAKE_CXX_FLAGS to the E2E tests can cause problems, for example a test might wish to override options such as setting -ffp-model=fast, but this fails with a warning/error if CMAKE_CXX_FLAGS contains -ffp-model=precise: error: overriding '-ffp-model=precise' option with '-ffp-model=fast' [-Werror,-Woverriding-option].

While the error could be worked around in the test by appending -Wno-overriding-option to the test flags, I don't believe this is the right solution.

It is possible that the user is running on a system where some flags have to be passed to clang++ to make it produce working executables. For this reason, allow setting SYCL_E2E_CLANG_CXX_FLAGS by the user if needed in in-tree builds, but do not pass CMAKE_CXX_FLAGS by default.

Contains a drive-by fix to make sure -Werror is passed to clang++ even when the c++ flags are overriden by the lit parameter --param cxx_flags=... (it can still be disabled using -Wno-error explicitly).

[Maetveis]

run_prebuilt_e2e_tests (Intel Battlemage Graphics, ["Linux", "bmg"]

********************
Failed Tests (21):
  SYCL :: Complex/sycl_complex_math_test.cpp
  SYCL :: Complex/sycl_complex_operator_test.cpp
  SYCL :: DeviceLib/imf/fp32_test.cpp
  SYCL :: DeviceLib/math_test_marray_vec.cpp
  SYCL :: DeviceLib/math_test_marray_vec_fp16.cpp
  SYCL :: Reduction/reduction_span_pack.cpp
  SYCL :: USM/copy.cpp
  SYCL :: USM/memops2d/copy2d_device_to_dhost.cpp
  SYCL :: USM/memops2d/copy2d_device_to_host.cpp
  SYCL :: USM/memops2d/copy2d_device_to_shared.cpp
  SYCL :: USM/memops2d/copy2d_dhost_to_device.cpp
  SYCL :: USM/memops2d/copy2d_host_to_device.cpp
  SYCL :: USM/memops2d/copy2d_host_to_shared.cpp
  SYCL :: USM/memops2d/copy2d_shared_to_device.cpp
  SYCL :: USM/memops2d/copy2d_shared_to_dhost.cpp
  SYCL :: USM/memops2d/copy2d_shared_to_host.cpp
  SYCL :: USM/memops2d/copy2d_shared_to_shared.cpp
  SYCL :: USM/memops2d/memcpy2d_device_to_device.cpp
  SYCL :: USM/memops2d/memcpy2d_device_to_host.cpp
  SYCL :: USM/memops2d/memcpy2d_shared_to_host.cpp
  SYCL :: USM/memops2d/memcpy2d_shared_to_shared.cpp

I believe the USM failures are unrelated, they are tracked by #19749

The other errors seem to be due to

level_zero backend failed with error: 20 (UR_RESULT_ERROR_DEVICE_LOST)

Which is possibly a driver problem? I'll update the branch and try again, but I don't expect these to be related to the changes here.

[Maetveis]

@intel/llvm-gatekeepers this is ready to merge :)

[sarnex]

nice find and fix!