Refactor RTS scan_through_elements_helper to struct

[danhoeflinger]

Here's what I propose (I'll make an issue):
In a separate PR, create a struct which encapsulates this in two layers:

1. struct scan_through_elements_helper- created once per kernel, with a few public entry points which have template arguments for the constexpr fields which change from invocation to invocation within a kernel (__init_present for instance). Pull in the 2 usages of dispatch as named functions of this struct, to allow it to be a private helper.

2. struct sub_group_scan - toggle on or off inclusive_scan at the struct level, with 2 public entry points for scan and partial_scan, each templated with __init_present.

I think it makes sense to wait until after #2657, since some template parameters go away with that PR.

Originally posted by @danhoeflinger in #2709 (comment)

[akukanov]

The part (1) is made irrelevant by #2830 (if/when accepted) where I take another direction for refactoring of scan_through_elements_helper by merging it into reduce and scan submitters - the helper operates quite differently in each case anyway, and many arguments of the helper are member fields of submitters.

The part (2) remains relevant.