Improve mem. ops. + NodeID-ElementID mapping bindings

[sergiorg-hpc]

The PR improves the copy of the returned ids when fetching the data from the reports, as well as the Python bindings related to the NodeID-ElementID mapping (#168):

• In the former case, the std::vector is swapped in a constant-time operation instead of a full copy. This is particularly useful when a large number of GIDs is requested.
• In the latter case, the binding of the NodeID-ElementID mapping from is improved, returning a numpy.array instead of a list of tuples. Hence, providing directly the result generated from the C++ code without any additional copies.

Using the same ~60TB report from #174 that contains 4.2M GIDs (1.5B compartments), the following table illustrates the performance obtained by fetching a single timestep and varying the number of GIDs requested:

 # GIDS |            v0.1.11 |                 PR
 ------- -------------------- -------------------
      1 |  0.010879278182983 |  0.001698255538940
     10 |  5.343438386917114 |  5.512796163558960
    100 |  5.943392753601074 |  5.604890823364258
   1000 |  6.406604766845703 |  6.062043905258179
  10000 |  6.451685428619385 |  6.049137830734253
 100000 |  7.101217269897461 |  6.462664365768433
1000000 | 12.762017250061035 |  9.955524444580078
4234929 | 27.868659019470215 | 18.158316850662230

On the other hand, when using in Python the function get_node_id_element_id_mapping() from #168, the following table shows the performance obtained varying the number of GIDs requested:

 # GIDS |            v0.1.11 |                 PR
 ------- -------------------- -------------------
      1 |    0.0158734321594 |  0.001111268997192
     10 |    2.6325545310974 |  2.509858608245849
    100 |    2.9563860893249 |  2.782548427581787
   1000 |    3.2792141437530 |  2.988783836364746
  10000 |    5.4664847850799 |  3.832525491714477
 100000 |   24.0672097206115 |  3.506523370742798
1000000 |  225.1766004562378 |  6.213826656341553
4234929 | 1017.6884262561798 | 11.914409875869751

Note that there has been other attempts to improve the copies of the voltages and even the tuples for the NodeID-ElementID mapping container, but the benefits turned out to be negligible compared to the source code complexity that was required.

[mgeplf]

it's funny, I had this, but the asm turned out the same in both cases; now I'm doubting that.

[sergiorg-hpc]

Thank you very much, @matz-e and @mgeplf!

[mgeplf]

Thanks for the change