[SYCL] Follow up fixes for group_sort extension

[againull]

• Change memory_required queries to return the exact size of local memory which needs to be allocated for group_sort algorithm, currently it is a bit more than required.
• Fix a bug in align_key_value_scratch where I didn't take into account that std::align changes value of KeysScratchSpace. To find scratch of the values I need to use original value of this variable.
• Fix mistake in the test where keys/values were mixed up. This is more of a cosmetic change.
  RunSortKeyValueOverGroup functions accepts two vectors - the first is considered keys, and the second is considered values.
  Verification is done in the same function (with the same assumption that the first vector is keys and the second is values).
  So it doesn't actually matter in which order aforementioned arguments are provided, test still serves it purpose.

[aelovikov-intel]

KeysScratchSpace bug seems to have been introduced in #14399, it would makes sense for the reviewer from that PR to review this one as he'd be more familiar with the surrounding code. Adding @dm-vodopyanov to reviewers.

[againull]

tag @andreyfe1

[aelovikov-intel]

Fix mistake in the test where keys/values were mixed up.

How did it pass before and how did you find it? Should we change some values for keys/values in the test as well?

[andreyfe1]

Thanks!

[againull]

Fix mistake in the test where keys/values were mixed up.

How did it pass before and how did you find it? Should we change some values for keys/values in the test as well?

This is more of a cosmetic change.
RunSortKeyValueOverGroup functions accepts two vectors - the first is considered keys, and the second is considered values.
Verification is done in the same function (with the same assumption that the first vector is keys and the second is values).

So it doesn't actually matter if we do:

RunSortKeyValueOverGroup<UseGroupT::WorkGroup, 1>(Q, Data, Keys,
                                                      Comparator);

or

RunSortKeyValueOverGroup<UseGroupT::WorkGroup, 1>(Q, Keys, Data,
                                                      Comparator);

Keys and Data vectors are filled with random numbers (and these two vectors are different) and test serves it purpose no matter the order of the aforementioned arguments, so additional changes are not necessary.

[aelovikov-intel]

This is more of a cosmetic change.

Please add this explanation to the PR description.

[againull]

This is more of a cosmetic change.

Please add this explanation to the PR description.

Added.

[aelovikov-intel]

LGTM for two out of three changes here. Still needs @dm-vodopyanov 's approval.

[dm-vodopyanov]

LGTM for two out of three changes here. Still needs @dm-vodopyanov 's approval.

Suggest to approve without my approval because code reviews for all previous group_sort patches were made by @andreyfe1, as Andrey is an expert in group_sort extension, and he put above his informal approval for the current patch.