This fixes all our GObject Introspection issues with GimpUnit which was
both an enum and an int-derived type of user-defined units *completing*
the enum values. GIR clearly didn't like this!
Now GimpUnit is a proper class and units are unique objects, allowing to
compare them with an identity test (i.e. `unit == gimp_unit_pixel ()`
tells us if unit is the pixel unit or not), which makes it easy to use,
just like with int, yet adding also methods, making for nicer
introspected API.
As an aside, this also fixes#10738, by having all the built-in units
retrievable even if libgimpbase had not been properly initialized with
gimp_base_init().
I haven't checked in details how GIR works to introspect, but it looks
like it loads the library to inspect and runs functions, hence
triggering some CRITICALS because virtual methods (supposed to be
initialized with gimp_base_init() run by libgimp) are not set. This new
code won't trigger any critical because the vtable method are now not
necessary, at least for all built-in units.
Note that GimpUnit is still in libgimpbase. It could have been moved to
libgimp in order to avoid any virtual method table (since we need to
keep core and libgimp side's units in sync, PDB is required), but too
many libgimpwidgets widgets were already using GimpUnit. And technically
most of GimpUnit logic doesn't require PDB (only the creation/sync
part). This is one of the reasons why user-created GimpUnit list is
handled and stored differently from other types of objects.
Globally this simplifies the code a lot too and we don't need separate
implementations of various utils for core and libgimp, which means less
prone to errors.
Functions creating a new GeglBuffer should trigger a warning if the
result if unused, because this is potentially a big memory leak.
Similarly objects created by functions creating new layers should be
handled (usually by adding the layer to the image with
gimp_image_insert_layer()), because they also come with a buffer and
possibly quite some important memory leak.
This function is not perfect and in particular doesn't seem usable with binding
because of GimpUnit being some weird mix between an enum and some kind of class.
So this will have to be fixed too. See #8900.
Now text layers are proper types, which means that the binding API will also be
nicer (e.g. `txt_layer.set_text('hello world')` in Python).
This commit also adds the param specs allowing to create plug-in procedures with
text layer parameters.
Finally it fixes the few calls in file-pdf-save (apparently the only plug-in
using specific text layer API right now) with explicit type conversion.
Jehan