[C++ pyclustering] Optimize parallel_for in case of short cycles

[annoviko]

Introduction
Current implementation is slower than PPL implementation in case of tiny arrays (see hsyncnet tests for more details). There is illustration:

for_loop_over_some_data {
    parallel_for(small_array) {
         some_work();
    }
}

Acceptance Criteria

• Optimization is required in order to have the same performance or higher.
• Apply the implementation for Windows implementation instead of PPL if it increase performance more.