We are more clearly separating opening files and reading from them.

Directory Tests was created in commit 40a1c5aa.

Reverting commit 8ef2473e. We no longer have preprocessor directives
since commit 2905904d.

The function would crash anyway because "extremum" is not in
`handler["readers"]` but the error messages would be less clear.

Useful to read shapefiles from conversion of Matlab data.

Avoid calling hunt in that case.

With `target_sources` rather than referring to their path from another
directory.

It was difficult to find the `CMakeLists.txt` that controlled the the
compilation of a given file from the message of a compilation error.

Since `NR_util` has been integrated into Jumble.

But reading from them is not here yet.

Not very important but cleaner.

Since `shpc_open` is called by all processes, reading d0 and
`ishape_last` in `shpc_open` implies that we do not broadcast d0 and
`ishape_last`.

Store `ishape_last` and first date in components of derived type
shpc. We then remove arguments `d_init` and `ishape_last` of procedures
`read_snapshot` and `get_snapshot`.

`read_field_indices` is now called from `shpc_open` so the comment
no longer makes sense.

It was confusing to mix opening an SHPC with broadcasting some part of
the resulting shpc variable.

Note that we then have to declare `d_init` before `ishape_last` in
procedure `get_snapshot`.

We are reverting commit e48b7f7e. I think I did not really appreciate
then that the number of days since 1950-1-1 is another date index. It
is clearer to use just one date index, which is in the dbf files. So,
in effect, we replace everywhere the old 0-based date index k by `k -
d_init`. Note that we no longer need argument `d_init` for procedure
`dispatch_snapshot` and we no longer need to write `d_init` to file
`node_id_param.json`. Note also that the value of every node index
changes by the amount `d_init * e_overestim`.

To search those names more easily.

When reading an eddy.