Also, remove the hacks that integrate Skia ref counting with WTF. Now we
just use sk_sp everywhere for managing Skia ref counts.

This patch introduces a //glue library that isolates our dependency on //base.
This will let us clean up our //base dependencies ahead of actually being able
to fully remove them.

This patch removes almost all //base dependency of //flow. The only dependency
left is on tracing.

I noticed this while doing the libjpeg_turbo roll.

* Added instrumentation for timings.

* A few more traces for better granularity.

Instead, make Layer::Paint take a new PaintContext that has just exactly
the state that it needs, mirroring PrerollContext. Also, rename
PaintContext to CompositorContext because it holds the context for the
whole compositor.

This patch makes these files better match Google C++ style.

Also, add the engine lap time when running on Mojo.

- fix dartdocs for addPerformanceOverlay
- make PerformanceOverlayLayer honour its x, y, and height.
- fix the y axis of PerformanceOverlayLayer to only show 3x16ms, since
  if it's more than 16ms it really doesn't matter what it is.
- make the label on the PerformanceOverlayLayer show the worst time on
  record not the instantaneous time.
- pin the fps to a maximum of 60Hz

Previously we just dropped child scenes on the floor. Now we upload them
to Mozart. However, we just draw them on top of all the other content
and don't apply any clips or blends.

Now that the compositor doesn't have any tricky deps, we can make it a
top-level project, which saves a bunch of typing.

Now we just depend on Skia and Base, which makes the compositor easier
to understand.

Also, update our skia/ext/refptr.h to match the current version of the
code in Chromium. This version adds support for several useful C++11
patterns.

Rather than relying upon the rects passed in from Dart, the compositor
should compute the paint bounds of layers from the cull rects of the
underlying SkPictures. This approach is better because it will handle
effects like shadows that paint outside the incorrect paint bounds we
use today (as well as shrinking around empty space).

Prerolling the rasterization tasks reduces the number of render target
switches because we don't interrupt the main render target.

The gr_context and canvas are stored at the frame level instead of the storing the same temporarily in the context

Make the paint context own the rasterizer and options. Also give it the ability to vend frames to render into.

dont mutate from frame to frame

We'll need to make this fancier as the compositor gets smarter.

Instead of squashing the layers down into a single SkPicture, we now pass the
tree to the GPU thread, which draws everything separately.

We now build a layer tree before squashing the draw command down to an
SkPicture.