XCB: Removal of cursor cache caused a big performance regression with many native windows

The patch in question:
https://codereview.qt-project.org/#/c/170426/

This leads to a very bad performance regression with X11 forwarding, tested around 10 times slower than with the patch reverted. Even locally I can see it increases time taken with the attached test case from 300ms (patch reverted) to 830ms.