cuda-parallel: Algorithms should not make repeated calls to get_bindings() function

[oleksandr-pavlyk]

In reduce_info algorithm the call to get_bindings() is made in __init__, __call__ and __del__ methods.

In merge_sort it is made in __init__, saved to self.bindings, reused in __call__ but made again in __del__.
Similar patterns repeats across algorithms.

This issue is to track work to only make the call to get_bindings() in __init__ method, and reuse self.bingings in all others across algorithms.