I had the directories `n_dates_1` and `n_dates_14`, but they did not
evolve with the code and there was not much traceability.

Dictionary instead of list, following update of `test_compare.py`.

This is much too long for completely different files.

Apart from the fact that `Global_1993_2021` was obtained with a more
recent version of the code, there is not much point in keeping a
separate experience for the year 1993.

Instead of symbolic links. We want to avoid creating symbolic links
above SHPC because we want to run several instances of `inst_eddies`
in parallel.

So the script can write to a slice other than 0.

Bug fix: do not provide standard input for `inst_eddies_Aviso.py`, it
does not read it.

Use out_cont%closed instead.

This avoids copying `cont_list(n_cont)` to `out_cont` if it has
insufficient area.

Add dummy arguments `cont_list` and `n_cont` to `get_1_outerm`.

Because we are going to add another intent out argument.

Remember that we do not write invalid eddies to shapefiles, so there
are no null shapes in `outermost_contour.shp`. So, in order to predict
the size of output, the interesting statistics is the number of points
divided by the number of non-null outermost contours.

Because the numbering changed when we decided not to write invalid eddies.

I do not remember why I chose to write to standard error. First it is
clearer to write only error messages to standard error. Second the
order of messages on standard error coming from the Fortran code and
from the shapelib C library does not seem deterministic, and this is
annoying to compare runs.

Because the numbering changed when we decided not to write invalid eddies.

Because `$tests_old_dir/Extraction_eddies_region_5/SHPC` is opened
with `SHPC_class` in test `Plot_eddy_contours_window`, which will
create `n_slices.txt`, so there will always be this file only in
`$tests_old_dir/Extraction_eddies_region_5/SHPC`.

We import in the test infrastructure the experiment `Inst_eddies_4`
because we want to no longer keep invalid contours, so we will need to
update the experiment.

As `grid_lon_lat` may be false, we define `e%extr%coord_proj`.

I do not see why this was not done in commit 24b5b079.

This shortens the file `test_get_1_outerm.f90`.

We will need this new shapefile in program `eddy_graph` only for a non
longitude-latitude grid. We add a new dummy argument `grid_lon_lat` to
procedure `shpc_create` to decide whether we create the shapefile
`extr_proj`.

Add component `innermost_level` to derived type eddy. So
`innermost_level` does not transit as an argument through main program
unit `inst_eddies` any longer.

Extract definition of extrema from procedure `set_all_outerm` into new
procedure `set_all_extr`. So we are reverting part of commit
f5f4298d. The drawback, for now, is that `innermost_level` has to
transit in the main program unit from `set_all_extr` to
`set_all_outerm`.

Rename file `test_write_eddy.f90` to `test_write_null.f90`. The name
was badly chosen because the program is not a unit test of procedure
`write_eddy`. In fact, it does not even call `write_eddy`.

Use relative path to Overlap from `Inst_eddies/Tests`. For symmetry
with what we do in `Overlap/Tests`. So now the directories Overlap and
`Inst_eddies` can be added from the top-level `CMakeLists.txt` in
either order.

Following update of NetCDF95.