[clang] [clang][SYCL] Diagnose variadic sycl_external function (PR #216393)

[Tom Honermann via cfe-commits]

================
@@ -624,6 +624,10 @@ following requirements:
 The attribute shall be present on the first declaration of a function and
 may optionally be present on subsequent declarations.
 
+The attribute is ignored, with a warning, when applied to a variadic function.
+A variadic function cannot be called from device code, so there is no reason to
+emit device code for one.
----------------
tahonermann wrote:

It's a little funny that we document this with a statement that "there is no reason to emit device code for one" and then have a test which validates that a variadic function is actually emitted if ODR-used (in a way other than a function call).

With the PR as is, we're arguably introducing a strange inconsistency by allowing a variadic function to be ODR-used if the ODR-use is in the same translation unit as the function, but not permitting such an ODR-use to occur from another translation unit (because ignoring `SYCL_EXTERNAL` prevents that).

I think we should proceed with the behavior in the PR for now. I filed a request with the Khronos Group to clarify what the intended behavior is. If and when the intended behavior is made clear, we can adjust the behavior accordingly.
- [SYCL-Docs #1061](https://github.com/KhronosGroup/SYCL-Docs/issues/1061): Can a variadic function be ODR-used by device code as long as it isn't called?

In the meantime, how about a tweak to note the special case?
```suggestion
SYCL prohibits calls to a variadic function in device functions. Since there is
little reason to emit a function that cannot be called, the attribute is ignored,
with a warning, when applied to a variadic function.
```

https://github.com/llvm/llvm-project/pull/216393