If an application sends a large number of events, then deferred_messages can
become quite large (possibly holding several thousand messages), so appending
to a GList can become a performance hit.  Thus, use a GQueue instead.

Attributes can now be cached, but not enabling by default, since there
is currently no event to notify AT-SPI that attributes have changed (see
BGO#649771), so this is dangerous but may improve performance if we can
reliably assume that attributes will not change.

Use the generated enum type to convert state names to numeric values rather
than a string table that needs to be maintained in parallel.  Also, remove
atspi_role_get_name since it was unused, marked deprecated, and relied on a
string table which is now redundant with the enum type.

Have atspi_accessible_get_role_name try to use the internal string associated
with the accessible's role rather than calling GetRoleName.  This avoids
making a DBus call when the role is cached.

Also, atspi.h now includes atspi-enum-types.h, and the latter is now installed.

Use glib-enums to generate GTypes for enums, and modify enum definitions so
that they will be marked as bitflags where appropriate.  This helps with
introspection and allows functions that take flags to be prototyped as such
rather than being marked as taking a gint.

If we are not in a GLib main loop, then we will not see signals to update
the cache, so cached data will become invalid.  Assume that, if
atspi-event_main has not been called, then we are, for instance, in a simple
pyatspi script without a main loop, so we should disable caching.

Putting reentrant call closures on the stack can be dangerous if set_reply
is called after the caller has timed out and returned, for instance, so better
to dynamically allocate them.  This will hopefully fix some crashes and
hopefully not leak memory.

Previously, atspi_accessible_set_cache_mask would not affect accessibles
which were already cached, but this was not very useful, as caching is done
automatically when a new app is discovered.  Functions that may retrieve
cached data now test the cache setting for the application--not just whether
there is a value stored for the data--whe ndetermining whether to call the app.

Also, annotate atspi_accessible_set_cache_mask as a gint.  This is temporary;
properly generating GTypes for enums/flags is TODO for 2.2.

It is not safe to ref a hash marked (transfer none) and expect its elements
to remain, since the caller will assume that it owns the elements and may
deallocate them (this is done by pygobject).  It would also be problematic
to simply mark the hash (transfer full), since we need to either assume that
the caller passed a destroy function when creating the hash (pygobject does
not) or deallocate the values ourselves when the match rule is finalized
(which requires making assumptions about the correct free function needed to
deallocate the values that the caller allocated).  So it seems that the only
safe thing to do is to copy the hash.

- atspi-application.h makes use of DBusConnection* .
- -Werror-implicit-function-declaration doesn't contain an equal sign in
  my copy of gcc, and is gcc specific.
- The default behaviour of AC_CHECK_LIB glib-2.0 is to add -lglib-2.0 to
  LIBS which can then cause the search for a dlopen function to fail
  because the flags to find libglib aren't included. Solution is to
  tell AC_CHECK_LIB to do nothing(!)

Previously, AT-SPI would wind up passing a NULL connection to libdbus if it
couldn't connect to the bus, causing a cryptic abort.  This change makes it
at least abort with a sensible error message in that case.
Also, if unable to find the accessibility bus address via the session bus,
save and print the DBusError, if any.

_atspi_dbus_get_property calls g_strdup when it returns a string, so we
should not g_strdup the result unless we're caching a copy for later use.

If malloc returns NULL, then g_malloc will call g_error and thus abort, so
it is redundant to check the return value of g_new0 and the like, since glib
will already have aborted the program rather than returning a NULL value.
"Cleaning up" the code by removing some of these checks.

The consensus seems to be that messages written out as a result of calls to
g_warning/g_log should ot be translated; ie:
http://mail.gnome.org/archives/desktop-devel-list/2011-March/msg00070.html

So removing translations from these calls.

When GetAPplicationBusAddress is called, at-spi sets a callback to handle the
reply.  It is possible that the application will go away and the
AtspiApplication will be disposed before this callback is reached, in which
case the reference to the accessibility bus connection will have gone away.
In this case, we should not call GetItems, since app->bus will be NULL and
libdbus will throw a SIGABRT.

This was causing occasional segfaults in the gdm session.
https://bugzilla.gnome.org/show_bug.cgi?id=648942