This issue came from the frame-based refactoring done in the multiwin
PR.
It looks like some (all?) compositors group certain events alongside
`wl_pointer::leave`, which I absolutely did not expect. The docs don't
seem to mention it either from what I can tell.
We now fall-back on the old pointed window if and only if the current
window is invalid and the old one isn't. Each state fetch is guarded
with an `ERR_FAIL_NULL` so this should still catch any potentially
corrupted window with missing data but a valid ID.
I also added the usual big comment so that this "quirk" does not get
lost to time.
a3913b045d