[SYCL] Fix types and transparent functors recognition in reduction

1. Enable operator*, operator+, operator|, operator&, operator^= for
   corresponding transparent functors used in reduction.
2. Fixed the case when reduction object is passed to parallel_for
   an R-value.
3. Allow identity-less constructors for reductions with transparent functors.
4. Replaced some 'auto' declarations with Reduction::result_type
   and added intermediate assignments/casts to avoid type ambiguities
   caused by using sycl::half type, and which may also be caused by custom/user
   types as well.

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

Author: v-klochkov

sycl/include/CL/sycl/handler.hpp

@@ -853,7 +853,7 @@ class __SYCL_EXPORT handler {
             int Dims, typename Reduction>
   detail::enable_if_t<Reduction::accessor_mode == access::mode::read_write &&
                       Reduction::has_fast_atomics>
-  parallel_for(nd_range<Dims> Range, Reduction &Redu, KernelType KernelFunc) {
+  parallel_for(nd_range<Dims> Range, Reduction Redu, KernelType KernelFunc) {
     if (Reduction::is_usm)
       Redu.associateWithHandler(*this);
     shared_ptr_class<detail::queue_impl> QueueCopy = MQueue;
@@ -886,7 +886,7 @@ class __SYCL_EXPORT handler {
             int Dims, typename Reduction>
   detail::enable_if_t<Reduction::accessor_mode == access::mode::discard_write &&
                       Reduction::has_fast_atomics>
-  parallel_for(nd_range<Dims> Range, Reduction &Redu, KernelType KernelFunc) {
+  parallel_for(nd_range<Dims> Range, Reduction Redu, KernelType KernelFunc) {
     shared_ptr_class<detail::queue_impl> QueueCopy = MQueue;
     auto RWAcc = Redu.getReadWriteScalarAcc(*this);
     intel::detail::reduCGFunc<KernelName>(*this, KernelFunc, Range, Redu,
@@ -920,7 +920,7 @@ class __SYCL_EXPORT handler {
   template <typename KernelName = detail::auto_name, typename KernelType,
             int Dims, typename Reduction>
   detail::enable_if_t<!Reduction::has_fast_atomics>
-  parallel_for(nd_range<Dims> Range, Reduction &Redu, KernelType KernelFunc) {
+  parallel_for(nd_range<Dims> Range, Reduction Redu, KernelType KernelFunc) {
     size_t NWorkGroups = Range.get_group_range().size();
 
     // This parallel_for() is lowered to the following sequence: