[SYCL] Don't define SYCL_EXTERNAL if -fno-sycl-rdc is passed

[sarnex]

Please review this commit by commit because one of the commits affects a ton of files. Thanks!

The SYCL spec allows us to declare we don't support device function calls outside of the current translation unit by not defining the SYCL_EXTERNAL macro.
With the -fno-sycl-rdc option, this is exactly that case: we do not support calls outside of the TU, so don't define the macro.

However, we define and use the SYCL_EXTERNAL in the SYCL headers also, so instead add __DPCPP_SYCL_EXTERNAL and use that. It is defined to SYCL_EXTERNAL if SYCL_EXTERNAL is defined, and otherwise __attribute__((sycl_device)) for device code or empty for host. This is fine because none of the uses here actually end up as cross-TU calls, it's either linked in from the device libraries or replaced by the SPIR-V translator to some SPIR-V intrinsic.

I added a test in sycl/test/basic_tests/macros_no_rdc.cpp to make sure nobody uses SYCL_EXTERNAL directly in the SYCL headers in the future.

(UPDATE: This part has been removed) However, with the macro being undefined, if a user tries to use it, they just get the classic unknown type name 'SYCL_EXTERNAL' which is not super helpful. I added some really gross code to give a better error message, but I'd like to hear from the CFE folks if there's a better way to do it, and if not, if it's so gross we should just drop it and rely on the normal error.

Relevant discussion: #8479

[sarnex]

 Failed Tests (1):
    SYCL :: SubGroup/shuffle_fp64.cpp

unrelated

[Fznamznon]

Since SYCL spec allows to not provide the macro, I wonder if it should be up to user to check whether the macro is defined by an implementation before using it?

[steffenlarsen]

Small question for my own education, but the runtime changes LGTM.

[elizabethandrews]

FE changes LGTM

[sarnex]

@v-klochkov Mind reviewing this? This is an NFC from ESIMD POV, just header macro name changes. Thanks!

[sarnex]

********************
Timed Out Tests (3):
  SYCL :: DeviceLib/assert-aot.cpp
  SYCL :: DeviceLib/assert.cpp
  SYCL :: KernelFusion/pointer_arg_function.cpp

unrelated

[sarnex]

@bader @intel/llvm-gatekeepers Hey all, can we merge? CI failures unrelated. Thanks!

[AlexeySachkov]

Since SYCL spec allows to not provide the macro, I wonder if it should be up to user to check whether the macro is defined by an implementation before using it?

It is indeed up to user to check if the macro is defined before using it