[NFCI][SYCL] Refactor reduction implementation

[aelovikov-intel]

The core part of it is adding the user's reduction variable type as an
explicit template parameter and modifying reduction to store just that,
instead of having three members (RWAcc/DWAcc/UsmPtr) of which only one
is used to store it.

That, in turn, made the whole default_reduction_algorithm redundant as
all it was used for was to propagate that exact information in an
obscure way.

Once done, it enabled some further simplifications:

• Simplifying some methods to use compile-time "if constexpr" instead
  of checking the type of the user reduction variable in runtime
  (i.e., previous "if (MRWAcc)" check).

• Making reduction_impl_algo::reduction_impl_algo ctor generic and
  having a single instance of it, eliminating three version for
  different types of user's reduction variable.

• Unify reduction_impl::reduction_impl ctors accessing RW/DW accessor
  eliminating code duplication

• create make_reduction helper function to make creation of
  reduction_impl require less boilerplate code by using template type
  parameters deduction (couldn't use CTAD as some of the params can't
  be deduced and there is no such thing as partial CTAD).

• Simplify reduction_impl creation further by benefiting from
  accessor's CTAD.

[aelovikov-intel]

Failures are unrelated, I believe:

Failed Tests (4):
SYCL :: dpas/dpas_test1.cpp
SYCL :: dpas/dpas_test2.cpp
SYCL :: dpas/dpas_test3.cpp
SYCL :: unused_spec_const.cpp

[steffenlarsen]

Minor comment, but otherwise LGTM.

[aelovikov-intel]

The failures are unrelated to this PR, I believe:

Failed Tests (4):
SYCL :: dpas/dpas_test1.cpp
SYCL :: dpas/dpas_test2.cpp
SYCL :: dpas/dpas_test3.cpp
SYCL :: unused_spec_const.cpp

Hence, @intel/llvm-gatekeepers this PR is ready.

[pvchupin]

Fails are getting addressed at intel/llvm-test-suite#1115 and intel/llvm-test-suite#1117