[SYCL] Pass handler & instead of queue across ABI for reduction utils

[aelovikov-intel]

Queue might be nullptr in case of graph, but the information these
utilitiess query is device-specific. By passing entire handler & and having
access to graph information we'd be able to return more precise results.

Another positive side-effect is that we eliminiate explicit
std::shared_ptr<queue_impl> which is a small step forward in the
ongoing refactoring efforts to prefer passing *_impl by raw ptr/ref
with explicit shared_from_this whenever lifetimes need to be extended.

[aelovikov-intel]

This is based on top of #18830, please ignore first commit during review. Also, while @intel/sycl-graphs-reviewers aren't technically owners of the reduction-related changes, it mostly affects the graph-related code path, so their review is still welcome.

[aelovikov-intel]

This is based on top of #18830, please ignore first commit during review. Also, while @intel/sycl-graphs-reviewers aren't technically owners of the reduction-related changes, it mostly affects the graph-related code path, so their review is still welcome.

Rebased, not relevant anymore. @intel/sycl-graphs-reviewers review request still stands :)

@slawekptak, this is ready for review.

[fabiomestre]

I think this is fine since graphs at the moment don't support reductions: https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc#12-non-implemented-features-and-known-issues