We still use SkPicture as our rasterization backend, but now we're uploading
our layer tree to C++. A future patch will push the layer tree deeper into the
system and we'll eventually switch backends.

We now use the repaint system to do all the painting. During initialization, we
set up a root layer that applies the device pixel ratio.

Fixes #706

The compositor backends we're planning to use can't handle a general-purpose
paint layer and instead need lower-level operations.

Fixes #707

- Introduce some new Layer classes.
- Introduce paintChildWith* methods.
- Convert paint() methods to use paintChildWith* where appropriate.
- Fix paintBounds logic in Layer world.
- Introduce Layer.replaceWith(), so that it's clearer what's going on.
- Make RenderObjects have a ContainerLayer, not a PictureLayer.
- Introduce a PaintingContext.replacingLayer() constructor to highlight
  where we are creating a layer just to replace an older one.
- Rename some layer-related methods and fields for clarity:
   requiresCompositing -> hasLayer
   hasCompositedDescendant -> needsCompositing
   updateCompositing -> updateCompositingBits
   _needsCompositingUpdate -> _needsCompositingBitsUpdate
   markNeedsCompositingUpdate -> markNeedsCompositingBitsUpdate
- After updating compositing bits, if we find that the bit changed, we
  now call markNeedsPaint().
- Reorder markNeedsPaint() logic for clarity.
- Make flushPaint() start at the deepest node.
- Make _compositeChild() avoid repainting children with hasLayer that
  aren't dirty, instead it just reuses their existing layer.
- Made RenderView reuse the RenderObject layer instead of having its own.
- Made RenderView have hasLayer set to true.
- Add various asserts and comments.

- Add Canvas.getSaveCount()
- Make RenderClipRect call context.paintChildWithClip instead of doing the clipping itself
- Make ClipLayer take a Rect instead of a Size
- Make PaintingContext.canvas read-only
- Add PaintingContext.paintChildWithClip()
- Minor rearrangings of code and style tweaks

This patch makes it easier to debug layerization issues.

Separating these notions makes them easier to work with because offset is
relative to the parent layer whereas size is intrinsic to the layer itself.
This patch fixes the underpainting bugs when compositing the stocks example.

Now we have the ability to draw multiple PictureLayers. We still squash all the
pictures together into a single SkPicture for simplicity. In the future, we'll
submit them to C++ separately and composite them on the GPU thread.

Currently we have a single PictureLayer that everyone draws into. A future
patch will teach the system to use multiple PictureLayers.