Instead of a single configure serial, which is simply not enough to mitigate event feedback loops.
This should (hopefully) address the recent d3d8:device / d3d9:device test failures on the Gitlab CI, without having to add more `flush_events` calls to the tests.
I believe it should also provide a more general fix to this race condition than the configure_serial mechanism.
-- v4: winex11: Also discard near-matching configs with similar sizes. winex11: Keep a list of pending window config requests.