[Spec change] Clarify "atomic_ref" with private address space

[sycl-issue-bot]

Please review whether KhronosGroup/SYCL-Docs#986 by @gmlueck requires any changes to the CTS.

If changes are required: Open a new PR addressing the changes and link it to this issue.
If no changes are required: Close this issue and proceed with the spec PR.

EDIT

Let's use this issue to track UB behavior in the CTS related to atomic accesses on private data. See @bader's comment below.

[gmlueck]

Since this is clarifying that something is UB, I think we cannot add a CTS test for this.

[bader]

@gmlueck, can we use this issue to tracking removing UB from the SYCL-CTS?
For example:

SYCL-CTS/tests/optional_kernel_features/kernel_features_separate_unit.cpp

Lines 42 to 54 in cea4ae6

	using AtomicRefT =
	sycl::atomic_ref<unsigned long long, sycl::memory_order::relaxed,
	sycl::memory_scope::device>;
	template <>
	SYCL_EXTERNAL void
	use_feature_function_external_decorated<AtomicRefT, sycl::aspect::atomic64>(
	const sycl::accessor<bool, 1>& acc) {
	unsigned long long temp = 42;
	AtomicRefT feature1(temp);
	AtomicRefT feature2(temp);
	feature1 += 42;
	acc[0] = (feature1 == feature2);
	}

[gmlueck]

OK, reopened and updated the description.