[KT][ESIMD Sort] Code quality improvements

[dmitriy-sobolev]

These are the gaps which remain after the main review (#1257):

1. Separate declaration and definition (can be postponed until productization):
   See discussion: [KT] ESIMD-based radix sort kernel templates #1257 (comment).
2. Use zip_iterator/zip_view instead of __rng_pack:
   See discusison: [KT] ESIMD-based radix sort kernel templates #1257 (comment) and [KT] ESIMD-based radix sort kernel templates #1257 (comment).
3. Consolidate trivial checks for the sequence length (see the checks like if (__keys_last - __keys_first < 2)).
4. Refactor __onesweep_memory_holder class.
   See discussion: [KT] ESIMD-based radix sort kernel templates #1257 (comment).
5. Consider renaming __vector_store overloads to denote that they work with SLM.
   See discussion: [KT] ESIMD-based radix sort kernel templates #1257 (comment).
6. Get rid of loops in __one_wg_kernel and use native functions or helper wrappers where possible. Avoid bitcasts to uint32_t.
   See discussion: [KT] ESIMD-based radix sort kernel templates #1257 (comment)
7. Make _ONEDPL_ESIMD_INLINE a genera macro to enforce inlining or replace it if there is an existing one.
   See discussion: [KT] ESIMD-based radix sort kernel templates #1257 (comment)