Complete the SSH sampling between `innermost_level_2` and
`outermost_contour`, and find associated contours in `set_max_speed`.

From `set_all_contours`, pass `n_cont` and the whole array `cont_list`
to `set_max_speed`, because we are going to define new elements into
this array in `set_max_speed`. `cont_list` in `set_max_speed` becomes
`intent(inout)`.

Because `cont_list` is going to become `intent(inout)`.

We keep ssh because we will need it if we call `good_contour`.

Remove dummy argument `ind_targ_extr` of procedure `set_max_speed`,
not used any longer.

Remove component radius4 of derived type eddy and field radius4 of
`outermost_contour.dbf`. They are not used any longer.

Abandon the radius4 approach, because the resulting max-speed contours
had speeds significantly lower than those from Matlab and because the
approach with radius4 is difficult to extend to a non
longitude-latitude grid.

Remove argument `outside_points` of `set_max_speed`.

Check we find a max-speed contour when `e%radius4 >= 2` and
`e%speed_cont%ssh /= missing_ssh`. See test `Set_max_speed_degenerate`
for a case when we do not find one.

Also split the tests, Fortran source files, Python files and JSon
files. Keep at the top level the files used in both `Inst_eddies` and
Overlap.

Motivations for the split:

- The top directory contained many files.

- It may be useful to compile only one of the two sub-projects. For
example, only `Inst_eddies` on Ciclad, where MPI 3 is not available.

Add options for the NAG compiler.

Replace `target_include_directories` by `include_directories` in
`Tests/CMakeLists.txt`. The file was becoming too long.

Remove unused items in use statements.

Bug fix in `Tests/Input/successive_overlap_region_5_nml.txt`: remove a comma.

spherical_polygon_area to spher_polygon_area.

In procedure max_speed_contour_ssh, instead of aborting when the speed
is not defined, continue with missing_ssh. Found such a case in
November 29th, 2015 data. Added corresponding test. In procedure
set_max_speed, if max_speed_contour_ssh returns missing_ssh then we
just take the outermost contour as max-speed contour.

after its initial computation and show more directly the consequence
of a NaN value. The drawback is the repetition of one line of code. As
for performance, one less test is performed in the case of a NaN
value.

In script stat.py, add stats for cases when speed is missing.

and debugging options, short tests each take less than 10 s, while
long tests each take about one mn.

Bug fix in plot_snapshot.py. Slicing a multiple-dimension array with a
scalar logical does not produce an array with the same shape.

Removed computations of number of points per contour from program
test_get_snapshot. Created instead a post-processing script, stat.py,
computing these numbers of points and other stats.

known bug since three years for plots that go back and forth across
the map longitudinal boundary.

values at longitude boundaries, in case of periodicity.

Add argument "periodic" to procedures get_snapshot and
set_all_outerm. Duplicate values at longitude boundaries in ssh, u, v
if periodic. set_all_outerm can now call local_extrema with periodic
true. In procedure set_all_outerm, we shift the longitudes of outside
points if necessary.

In program test_get_snapshot, define periodic from step and nlon.

Stdin to Input. Two input dir provided for in tests.json: $input_dir
and $large_input_dir.

Rename everywhere variable nearby_extr to outside_points to avoid
confusion with function nearby_extr. Rename file "nearby_extr.csv" to
"outside_points.csv".

In procedure nearby_extr, include in the selection extrema with
different sign, even if they do not have sufficient amplitude.

Bug fix in read_eddy: use the value of read attribute to define
e%interpolated.

procedure get_snapshot. The idea is that, in this branch, nearby_extr
will eventually include extrema depending not only on them having
sufficient amplitude but also on their cyclonicity. Both pieces of
information cannot be simply encoded in extr_map. The downside is that
we compute needlessly nearby_extr if e%radius4 == 1 in set_max_speed.

set_max_speed. Abandon the computation with pack_indices, this was
supposed to give the same result than the computation with argwhere
but be faster if not any(extr_map) < 0. I was not sure of that. I
abandon that because I want to eventually remove the coding of
sufficient amplitude in extr_map.

instead of defining the component ssh to the ssh of the extremum or to
the ssh of the outermost contour, use a missing value flag
(1e4). Also, when radius4 >= 2 and the max speed contour is the
outermost contour, use missing value flag in the ssh component of max
speed contour instead of ssh on outermost contour. This makes clearer
output.

Add function null_ssh_contour.

In procedure get_1_outerm, simplify the comparison of level_try and
level_good. So we do not need any longer the dummy argument
ssh_extremum.

of local variable radius in procedure set_max_speed. Add corresponding
column in outermost_contour DBF file. This is just a diagnostic of the
workings of get_snapshot. Clearer to have this output than having
special meaning in components ssh and area of speed_cont.

weight. So, in procedure set_max_speed, in case radius equals 1, set
e%max_speed to the speed on the outermost contour instead of zero.

In procedure set_max_speed, in case the highest speed is on the
outermost contour, set e%speed_cont%polyline to null_polyline instead
of e%out_cont%polyline. This saves main memory and there is less
copying to do.

speed_cont can now be null even if max_speed is not zero so change
test for null shape in procedure write_eddy.

It appears that the shapelib C library cannot write a NaN value to a
dbf file, so write 1e4 instead.

Bug fix in plot_snapshot.py.