This seems to be a little faster, even if we do not use out-degree for
all the nodes. For `Global_1993_2023/Graph_anti`, this commit reduces the time
to define `traj_prop` from 23 s to 20 s.

We know that we are going to need in-degree for every node of the
graph. For `Global_1993_2023/Graph_anti`, this commit reduces the time
to define `traj_prop` from 31 s to 25 s.

This looked good but was much longer. For example, for
`Global_1993_2023/Graph_anti`, this commit reduces elapsed time
approximately from 24 mn to 2 mn.

This makes quite a difference in file size for a large domain. For
example, the size of `Global_1993_2023/Graph_anti/expanded_traj.json`
is reduced from 0.6 to 0.3 GiB. We do not need pretty display for
these files.

This is convenient for scripts `plot_traj.py`, `filter_traj.py` and
for hacking with function `get_duration` of `plot_traj.py`.

As an exception to our general philosophy, we prefer here to use as
input the results from previous tests in the current series of tests
because these are performance tests that we want to try with different
sizes of edgelist. We do not want to run the tests three times to have
the performance for a new size of edgelist.

So we can better distinguish initial position from small islands on
the map.

This is case 2 correction in the Matlab program of Rémi Laxenaire.

The test is useless in this version since we know `traj_prop[n] == -
1` at the beginning of the iteration, but we are preparing to process
phantom patterns.

Preparing to process phantom patterns.

Define `traj_segm` and `expanded_traj` after the loop on
vertices. This is clearer, and we are preparing to process phantom
patterns.

Preparing to define `traj_segm` and `expanded_traj` out of the loop on
vertices.

Change type of vertex property name in graph-tool graphs from string
to int. The option used in function `load_graph_from_csv` requires
version >= 2.49 of graph-tool.

Check output file suffix at the beginning of the script. No sense
going through the calculation if the suffix is wrong.

The old algorithm did not give correct results on tests
`Greece_trajectories`, `Test_order_edges` and `Traj_component`. We
abandon the idea of processing edges in date order. Instead, we now
process vertices in topological order.

`my_subgraphs` is only used inside the test on `"cost_function"`.

This can happen if the input graph file contains only an extracted
component.