ATK_STATE_COLLAPSED is not actually a macro, so replace #ifdef with a
version check, since we know it appeared in atk 2.38.

Only use ATK_STATE_COLLAPSED if atk is new enough to support it,
and lower the atk requirement in meson.build.
Atk needs a release, so otherwise we depend on unreleased git code.

This is required if for some reason the AT_SPI_BUS property is not set.
That happens for instance if for some reason the dbus session bus (and thus
the at-spi bus) is reused between X sessions.

4ebe0720 ("Fixed bug 592987 - Java Combobox is not accessible")
made the collapsed java state mapped to ATK_STATE_EXPANDED instead of
ATK_STATE_DEFUNCT, then fd325d7d ("Fixed bug 593261 - Slider bar can
not be controled by gok") turned it to ATK_STATE_INVALID, but we do have
ATK_STATE_COLLAPSED which really does seem to match the collapsed java
state.

OpenJDK seems to be sending flurries of visible data changed events, which
overloads us. They are however usually just for the same object, so we can
compact them: there is no need to queue another one if the previous hasn't
even been sent!

because (0,0) is a valid position and Orca would happily use it

Details of role/states retrieval, and getting the jni env, is too verbose.
We however like to see the object name and role

Giuseppe Capaldo authored Aug 01, 2019 and Samuel Thibault committed Aug 01, 2019

The TableCell interface is to be presented by the children, not the
table.

They do not actually properly respect start/end boundaries