[SYCL] Do additional mostly NFC changes for reduction patch(1585)

[v-klochkov]

• Moved 2 big functions implementing command group functions
  for reductions from handler.hpp to reduction.hpp
• Removed handler::dissociateWithHandler()
• Removed handler::addEventToQueue() and made queue_impl::addEvent() private again;
• Minor changes in comments.
• Replaced 'auto' with 'size_t' in couple places.

[v-klochkov]

Vlad, please review this patch.
It was done for your comments in #1585

This patch did not yet include changes for your comment: #1585 (comment)

The main idea of that comment is to make the code look more graceful, in particular in finalize() method, if I understood it right.

Making the handler for which parallel_for(nd_range, reduction, func) called) empty and aggregative will also require fixing some other places that now require that each command-group-function must have at least one action/task (like single_task/parallel_for/copy). The alternative to that is to change the MCGType for 'this' handler from detail::CG::KERNEL to some new type.
Taking that into account makes that approach less graceful and thus compromises the whole idea.

I think this deserves a separate discussion and a patch sometime a bit later.

[v-klochkov]

Currently, there are 2 patches to simplify the review process.
1st patch simply moves 2 functions from handler.hpp to reduction.hpp
2nd patch addresses other reviewer's comments in #1585

[romanovvlad]

patch did not yet include changes for your com

Vlad, please review this patch.
It was done for your comments in #1585

This patch did not yet include changes for your comment: #1585 (comment)

The main idea of that comment is to make the code look more graceful, in particular in finalize() method, if I understood it right.

Making the handler for which parallel_for(nd_range, reduction, func) called) empty and aggregative will also require fixing some other places that now require that each command-group-function must have at least one action/task (like single_task/parallel_for/copy). The alternative to that is to change the MCGType for 'this' handler from detail::CG::KERNEL to some new type.
Taking that into account makes that approach less graceful and thus compromises the whole idea.

I think this deserves a separate discussion and a patch sometime a bit later.

Yes, my idea was to add a new type for MCGType.