Compatibility: Only use sycl::half if SYCL_CTS_ENABLE_HALF_TESTS is set

[fknorr]

This PR improves compatibility with implementations that do not expose the sycl::half type (notably, SimSYCL built with GCC).

There might be additional usages I missed, in tests that SimSYCL is currently unable to compile for other reasons.

Question to the reviewers: Is it necessary to exclude some translation units in CMake when the file names are *fp16.cpp or are these handled automatically?

[AlexeySachkov]

Question to the reviewers: Is it necessary to exclude some translation units in CMake when the file names are *fp16.cpp or are these handled automatically?

From what I see, they are filtered out automatically:

SYCL-CTS/tests/CMakeLists.txt

Lines 162 to 164 in 42e9081

	if(NOT SYCL_CTS_ENABLE_HALF_TESTS)
	list(FILTER test_cases_list EXCLUDE REGEX .*_fp16\\.cpp$)
	endif()

[keryell]

g++-13 implements already https://en.cppreference.com/w/cpp/types/floating-point from C++23. @fknorr Go for it! :-)

[keryell]

@fknorr Some merge from upstream?

[fknorr]

I was not aware of the _Float16 support in newer GCC - this PR might not be required anymore for SimSYCL, I can check after merging #870 .

[keryell]

@fknorr #870 has been merged and Ubuntu 24.04 comes with g++-14.
The more we wait, the more the features. 😸

[fknorr]

Sorry for abandoning this for so long!

I've rebased on top of #874 to avoid future merge conflicts.

@keryell unfortunately we still need the #ifdef chain since the CTS uses cmath functions (std::signbit etc) on all scalar float types, and these are not implemented for _Float16, at least not with libstdc++. We could get around this by raising the minimum C++ standard to 23 for SimSYCL (because 23 has std::float16_t including math builtins), but I'd rather not do this for compatibility reasons.

Edit: DPC++ passes these - but how?

[fknorr]

@bader's suggestion on reordering the functions in math_reference.h turned out to be more painful than expected, since moving function prototypes around messed with overload resolution in DPC++ (not sure if the code was actually incorrect or if I was triggering a compiler bug).

I've gone ahead and re-grouped all definitions within the file according to feature macros:

1. All scalar float and int functions
2. Scalar half functions (#if SYCL_CTS_ENABLE_HALF_TESTS)
3. Scalar double functions (#if SYCL_CTS_ENABLE_DOUBLE_TESTS)
4. All vec overloads (referencing scalar versions)
5. All marray overloads (#if !SYCL_CTS_COMPILING_WITH_ADAPTIVECPP)
6. Overloads generic over both scalar type and dimensionality.

A few functions (like cross) did not fit the schema above, so I kept those with their scalar cousins.