Rewrite following an idea of R\'emi Laxenaire: loop on edges in order of
ascending cost function.

Rename trajectories to `traj_segm` and `expanded_trajectories` to
`expanded_traj`.

No need for this property to be internal.

Instead of a global variable with hidden modification in a method of Visitor.

So that segments in a trajectory are in chronological order.

This is a new reversion in a series of reversions. See commit
836392d7. The idea now is to allow us to run a single test in a series,
taking advantage of the archive option in recent versions of
`test_compare.py`.

We are in the else branch of `if src.out_degree() > 1` and src is the
source of an edge so `src.out_degree()` must be 1.

Correct name of property from `segment` to `inst_eddies`.

Correct name of property from `nl_cost_function` to `cost_function`.

The maximum speed contour may not be defined separately from the
outermost contour. So if the radius from the maximum speed contour
shapefile is not defined, use the radius from the outermost contour
shapefile.

Motivation: performance. We eliminate repeated reading of the same
information from the shapefiles. If the length of the segment is
smaller than 14 days then we re-use for the last part of the segment
some of the properties from the first part. Also, we do not compute
properties for the first part of the segment if the segment has no
predecessor and for the last part of the segment if the segment has no
successor.

Remove intermediary variables `avg_rad` and `avg_Rossby`.

We extract part of function `calculate_radii_rossby` into new function
`node_to_prop`. So we now separate reading data from the shapefiles
and computing the average radius and Rossby number. Thus, we prepare
for rewriting the algorithm in the main part of the script. The
quantity of reading is still the same, for now.