ViewRoot: More surgically fix child-life time.

A more targeted version of "ViewRootImpl: Fix child lifetime".
In this version we avoid hijacking the window visibility callback
and introduce a new path.
Our contract has been that at any time after returning from
onStop, the WindowManager may destroy the clients Surface. We
see in setWindowStopped, we set the stopped state on the hardware
renderer in order to prevent further use of the Surface. However
there is nothing synchronously dispatching this change to child views
and so SurfaceView is not necessarily informed to release it's Surface
in time. Typically SurfaceView will be informed through the
onWindowVisibilityChanged callback. The signal producing this
is emitted from SystemServer prior to onStop but has to pass
through the clients handler thread. It seems this has always been
broken, and we have always had intermittent reports of EGL_BAD_ALLOC
scenarios. I speculate that elevation of WindowManager scheduling during
app close scenarios is perhaps making this more severe, as it seems to
ensure the WindowManger will win the race to destroy the Surface before
the client handler thread process onWindowVisibilityChanged (unless
the FIFO timeout is surpassed).

Test: Put Chrome in PiP. Turn screen off. No Crash!
Bug: 36561071
Change-Id: I9233ba8151c7f577b1d1044afc0c2ac3a65be4b4