Created attachment 241025 [details] [review]
giscanner: use SqlAlchemy's OrderedDict implementation

g-ir-scanner can be a bit on the slow side. While true we
now do a bit more work parsing GTK-Doc comment blocks and
more is still to come, one of the biggest hotspots besides
what's going on in _giscanner.SourceScanner() comes from
the OrderedDict implementations we have been using.

For example, time needed to build Gtk-3.0.gir on a relatively
slow machine using "python2 -m cProfile -o $prefix/bin/g-ir-scanner ...":

1) Our original DictMixin sublass:
     92,79867 seconds
2) Python's collections.OrderedDict class:
     88,65786 seconds
3) Larosa/Foord implementation from http://www.voidspace.org.uk/python/odict.html :
     71,64323 seconds
4) SqlAlchemy's implementation:
     66,12449 seconds

Looks like we have a clear winner with the SqlAclchemy
implementation, which comes in at around 20 seconds
without profiling on the same machine. Not bad.