Previously, only read-only layouts skipped same-layout transition.  This
is extended to color attachment and depth/stencil attachment layouts as
well.  On Nvidia+Linux, this does reduce the number of barriers in the
Draw* benchmarks, though there's no visible CPU-side performance
difference.  GPU time is not measured in these tests, but it's possible
that the driver already skips such transitions (i.e. this only saves us
from recording an ineffective transition).

Additionally, transfer dst and shader write same-layout transitions are
turned into an execution barrier instead of a memory barrier to avoid an
unnecessary flush.  Currently, this particularly affects texture upload,
and shows a 10% improvement in TextureUploadSubImageBenchmark.  Note
that this optimization is temporarily disabled due to a bug in the
windows AMD driver.

Bug: angleproject:3347
Change-Id: I7dc9d0b5dd2ad87ec19ae13277b330438038519f
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1659149Reviewed-by: Tobin Ehlis <tobine@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>