[KT][ESIMD sort] Potential performance improvements

[dmitriy-sobolev]

These are the gaps which remain after the main review (https://github.com/oneapi-src/oneDPL/pull/1257):

1. Use global block load functions instead of gather when possible. See discussion: https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1414570090 and https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1422767127
2. Shrink types: variables and loop counters. See discussions: https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1422781804 and https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1415632134
3. Replace loops by vector operations. See discussion: https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1415645085
4. Consider __data_per_work_item to be 32 (if this is a minimal granularity of data_per_workitem) See discussion: https://github.com/oneapi-src/oneDPL/pull/1257#discussion_r1415849750