[SYCL] Do not use handler::addAccessorReq()

[Alexandr-Konovalov]

This function might be useful only when an accessor is connected with different
CG that was used during a construction of the accessor. Currently it’s helpful
only in single case of reduction, and in this case the accessor can be connected
to a right CG in the constructor. Modify the use case and remove all calls to
addAccessorReq().

[Alexandr-Konovalov]

@intel/llvm-gatekeepers Could you please look at the PR?

[steffenlarsen]

My biggest worry about this patch is that we no longer keep accessors alive in the accessor storage of the command-group handler for many of these paths. Is this safe? Are we confident there are no scenarios that need them alive? I can see the graph copies the storage, @intel/sycl-graphs-reviewers.

[Bensuo]

My biggest worry about this patch is that we no longer keep accessors alive in the accessor storage of the command-group handler for many of these paths. Is this safe? Are we confident there are no scenarios that need them alive? I can see the graph copies the storage, @intel/sycl-graphs-reviewers.

I think this change should be fine for us in general if I'm understanding it correctly. Accessors are by default stored when associated with a handler (either through the ctor taking a handler or handler::require()) via this code. So outside of this niche usecase where the accessors needed to be associated with an additional handler I think this would have just been creating duplicate requirements for calls that took these paths.

We don't currently support reductions in graphs but I don't think this change makes much difference to the challenges of us implementing support for them so it seems fine in that regard also.

[steffenlarsen]

Thanks, @Bensuo! In that case, I forfeit my concerns. Let's roll with it!