core: introduce sample sort algorithm

The algorithm runs sort in multiple threads and does not use OpenMP. It
has better threads utilization right from the beginning but probably
a worse constant than parallel qsort. See details in code comments.

Besides sort is not performed in calling thread but instead in spawned
worker threads. Calling thread yields waiting for worker threads to
finish. Exception is small data size, in this case sorting is executed
in calling thread saving time on spawning a thread. This should speed up
test execution. This is existing behaviour of qsort_arg but data size
threshold is reduced from 128000 to 1024.

Part of #3389

NO_CHANGELOG=internal
NO_DOC=internal