Do more internal refactoring in reduce-then-scan

[akukanov]

Changes in this PR:

• parallel_backend_sycl.h - moved blocks of code without modifications for better grouping within the file.
• utils_ranges_sycl.h - removed unnecessary template parameter in the implementation of __require_access
• moved block carry routines out of scan submitter (because I will later need a similar routine in both scan and reduce submitters).
• moved __temp_data_required next to the input generators, which this trait is applied to.
• changed the scan-through-elements helper stack:
  • lowered communication tag dispatching to the "_impl" routine.
  • made the reduce submitter use "_impl" directly; the respective code path in __scan_through_elements_helper was trivial.
  • moved the functionality of __scan_through_elements_helper into a member function of the scan submitter, and merged more of operator() code into this new routine, including processing of out-of-bound writes and stop positions

The changes reduce the overall amount of code, set a single condition (__temp_data_required) for both reads and writes (instead of a different and seemingly unrelated __has_final_pos for writes), and reduce the API surface of scan-through-elements routines via direct access to data members of scan submitter.