For PixelUnpack, if the feature is enabled, it always create a shadow
CPU mapped memory. Then when mapBuffer is called, it let user write to
the shadow memory. At unmapBuffer time, it memcpy the shadow buffer data
to context's staging buffer and then issue a vkCmdCopyBuffer from
context staging buffer to the PixelUnpack buffer. This involves too
many data copies. The proper way to do it is simply make sure to create
the buffer as CPU map-able and let user directly write to it. If we find
cases that CPU waiting for GPU to finish is causing performance issues,
there are two ways to improve/fix it: 1)Add event at end of each
renderpass or FBO so that we know if GPU access is finished or not
without have to wait for entire frame to finish rendering. 2) Create
multiple buffers and ping-pong between them, at least there will be no
VkCmdCopyBuffer call involved like shadow buffers do. But we will leave
that to future time when we find such need.

Bug: angleproject:5986
Change-Id: Ib8300e46e779d20533c1f7f81624de0ce003788b
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2909758
Commit-Queue: Charlie Lao <cclao@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>