In my contrived test case here, this gets object inflation overhead down
by about 50%.

Personally, if I were doing this all over again from scratch, I would go
in a very different direction. But lets see how much better we can make
this without having to resort to a new ABI.

This is helpful so that we don't see all those deprecated warnings
everywhere. Yeah we should move away from those when we bump GLib
requirement, but no not now.

Realistically we could also shove this in private class data but might as
well avoid that for now.

Instead use a union that lets us switch bewteen the GParameter array and
the item resource. Keep track of how many parameters we have in there too.

Optimize the inflation of groups by finding the mappings once and then
using that instead of column iteration. Avoid a lot of GTypeClass ref'ing
as well. Only look up the quark data once per mapped parameter too.

We keep an extra GParameter at the end so that we can avoid copying the
GParameter for inflation somewhere else. We just add the repository at
the end of it before constructing/replacing the inflated instance.

Weak pointers are *extremely* slow in GObject and we don't really need to
use that anyway. We really don't have a good multi-threaded story here,
but I'm using a GMutex around the repository just to try to keep things
up with whatever they were before naturally in add_weak_pointer(). There is
a lot of unsafe things that can happen once you start threading anyway.

This release contains numerous performance improvements based on usage
from 0.5.0.