[SYCL] Implement parallel_for(range, reduction, func) (#4101)

* [SYCL] Implement parallel_for(range, reduction, func)

Currently parallel_for accepting sycl::range may handle only 1 reduction
variable.

Also, this patch had to change/update the methods Reducer::atomic_combine(),
which was the good moment to switch the methods implementation from
the deprecated sycl::atomic class to sycl::ONEAPI::atomic_ref.
The conditions on which the fast-atomics implementations are used were
not changed as that deserves a separate patch.

parallel_for accepting sycl::range works much faster than parallel_for
accepting sycl::nd_range, which means that nd_range version needs some
additional performance tuning soon.

Signed-off-by: Vyacheslav N Klochkov <[email protected]>

Author: v-klochkov