Replace contiguous NrnThread->_data with smaller memory chunks used in the CoreNEURON

[pramodk]

Currently, NEURON constructs NrnThread for CoreNEURON with one big memory chunk via nt->_data:

    nt._data = (double*)ecalloc_align(nt._ndata, sizeof(double));
    nt._actual_rhs = nt._data + 0 * ne;
    nt._actual_d = nt._data + 1 * ne;
    nt._actual_a = nt._data + 2 * ne;
    nt._actual_b = nt._data + 3 * ne;
    nt._actual_v = nt._data + 4 * ne;
    nt._actual_area = nt._data + 5 * ne;
    nt._actual_diam = ndiam ? nt._data + 6 * ne : nullptr;
    for (auto tml = nt.tml; tml; tml = tml->next) {
        Memb_list* ml = tml->ml;
        ml->data = nt._data + (ml->data - (double*)0);
    }
....
    int extra_nv = (&nt == nrn_threads) ? nrn_extra_thread0_vdata : 0;
    if (nt._nvdata + extra_nv)
        nt._vdata = (void**)ecalloc_align(nt._nvdata + extra_nv, sizeof(void*));

IIRC, there are also vdata members that can have global offsets to nt._data.

I believe it will be helpful if we can change these global offsets with more local ones and remove one big memory block nt->_data with smaller ones because of the following reasons:

• if we have to allocate different type of memories for different mechanisms or mechanism properties, currently it's not possible. This we require for GPUs or even KNLs.
• if we decide perform some of the mechanisms on GPUs on some on CPU, it's difficult to do because we have to copy all data to GPU (because of global memory offset).
• when we have to construct data structure piece-by-piece or use some nice C++ containers instead of raw pointers, it's not possible.
• @alkino is performing some of the refactoring of how we load the data in nrn_setup Split nrn_setup.cpp in phase1 / phase2 BlueBrain/CoreNeuron#283 but these global offsets and contiguous memory requirement pose a significant challenge in simplifying the code.

@nrnhines : we discussed this in the past but never able to make this priority. As Nico is doing very nice work on simplifying coreneuron code in BlueBrain/CoreNeuron#283, do you think we can make this change? This will also help NMODL and how we generate the code.

[nrnhines]

This is a very useful project. Past transformations involved the possibility of SoA memory layout and data permutation for higher performance ordering on gpu. Things have likely become so complex at
this point that a complete review is in order. The aspect that currently relies on a single double array
for everything is the use of indices instead of pointers into that array. For example, one of the more recent changes is communication of variables (trajectories and single values) between NEURON and CoreNEURON to allow vector recording and graphics. The key transformation on the
NEURON side is nrn_dblpnt2nrncore in src/nrniv/nrnbbcore_write.cpp and
double* stdindex2ptr in coreneuron/nrniv/nrnsetup.cpp on the CoreNEURON side. Clearly there needs to be some structure more complex than a mere int to replace pointers on the CoreNEURON side.