Design a new memory allocation scheme for derived variables

[anagainaru]

Current implementation

In the current release (https://github.com/ornladios/ADIOS2/releases/tag/v2.10.1) memory for a derived variable is allocated each time the variable values are computed.

template <class T>
T *ApplyOneToOne(T *data, size_t dataSize, size_t nVariables, std::function<T(T, T)> compFct)
{
    T *outValues = (T *)malloc(dataSize * sizeof(T));
}

For a complex math expression, each intermediate data is stored in a newly allocated buffer and freed at the end.

E.g.

            +
          /   \
         a    *
             /   \
            b   c

will allocate memory for temp = b*c and will allocate memory for a + temp

Future optimizations

We need to use a pre-allocated buffer for derived variables in the same way as we do with normal variable buffers and use parts of it for temporary data.

#4253 is removing complex numbers from derived variables because of memory allocation issues (this needs to be reintroduced once we implement a more elegant solution of dealing with memory)