Testing shows most app traces seeing performance improved on ARM GPU
when we flush at FBO boundary instead of one submit per frame. This
likely due to more frequent submission means less chance of GPU
starvation. This also helps ANGLE to detect GPU finish at finer grain
since the queue serial is per submission. Finer grain detection of
object completion means memory gets recycled quicker and mapBufferRange
able to hit fast code path more often. With this CL, manhattan 3.1
offscreen score also score improves 7%. This may also helps onscreen
performance if frame time is on the edge of 16ms by submit FBO rendering
earlier instead of been blocked by vkAcquireNextImage.

Bug: b/187993625
Bug: angleproject:5295
Change-Id: I2946a7159231d36019ace89805a69ac8e906ce94
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2888709Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Ian Elliott <ianelliott@google.com>
Commit-Queue: Charlie Lao <cclao@google.com>