ndarray size optimization

[elius1]

If an empty allocator is used, the allocator will still add sizeof(void*) to sizeof(vt::ndarray). This could be optimized away by using a std::tuple for the members thanks to empty base optimization.

Note: Check if this does not introduce performance regressions.

C++20 alternative that is much cleaner: https://en.cppreference.com/w/cpp/language/attributes/no_unique_address