[SYCL][ESIMD] Fix local accessor scatter test failure on PVC

[fineg74]

No description provided.

[v-klochkov]

The code "before" looks correct, and the new code is potentially UB, because many threads write to the same memory.

[v-klochkov]

Ok, I see now that Threads = 1, which makes this code correct. But it looks very confusing.
If look only to this code (and forget about Threads being set to 1), then this code looks like UB

[v-klochkov]

If there is some error in GPU Driver, then let's have Threads = 1 for now and keep this code, but please add a TODO comment here explaining that if Thread > 1, then this code needs an initialization done by only 1 thread and need a barrier between SLM initialization and the usage of SLM by all threads of the work-group.

[fineg74]

Restored the previous logic.
Working on a simpler reproducer to nail down GPU issue

[fineg74]

Windows E2E test failures:
SYCL :: Assert/assert_in_kernels_win.cpp
SYCL :: Assert/assert_in_multiple_tus_one_ndebug_win.cpp
SYCL :: Assert/assert_in_multiple_tus_win.cpp
SYCL :: Assert/assert_in_one_kernel_win.cpp
SYCL :: Assert/assert_in_simultaneous_kernels_win.cpp
SYCL :: Assert/assert_in_simultaneously_multiple_tus.cpp
SYCL :: Assert/assert_in_simultaneously_multiple_tus_one_ndebug.cpp
SYCL :: Plugin/sycl-ls-unified-runtime.cpp

are not related to the change