To make it easier to avoid pushing twice in one frame, provide a
transaction mechanism for the navigator.

This package contains WidgetTester, which is very useful when writing tests for
widgets.

- force the time dilation to 1.0 for the Widget tests, so that a local
  change doesn't break all the tests during development.
- add missing license block to all the files.
- set ui.window.onBeginFrame to null when you use WidgetTester, so that
  the engine doesn't trigger any confusing frames after our fake frames.

- Change RouteArguments to pass the route's BuildContext rather than
  the Navigator. This caused the bulk of the examples/ and .../test/
  changes (those are mostly mechanical changes). It also meant I could
  simplify Navigator.of().

- Make initState() actually get called when the State's Element is in
  the tree, so you can use Foo.of() functions there. Added a test for
  this also.

- Provide a RouteWidget so that routes have a position in the Widget
  tree. The bulk of the route logic is still in a longer-lived Route
  object for now.

- Make Route.setState() only rebuild the actual route, not the whole
  navigator.

- Provided a Route.of().

- Provided a Route.writeState / Route.readState API that tries to
  identify the clients by their runtimeType, their key, and their
  ancestors keys, up to the nearest ancestor with a GlobalKey.

- Made scrollables hook into this API to track state. Added a test to
  make sure this works.

- Fix the debug output of GestureDetector and the hashCode of
  MixedViewport.

- Fixed ScrollableWidgetListState<T> to handle infinite lists.

(These are changes cherry-picked from in-flight branches since they are
more independent and could be helpful even without those changes.)

- Change RouteBuilder's signature to take a single argument in which the
  other fields are placed, so that we can keep iterating on those
  arguments without having to break compatibility each time. Also, this
  makes defining route builders much simpler (only one argument to
  ignore rather than a variable number).

- Expose the next performance to RouteBuilders, since sometimes the
  route itself might not be where it's used.

- Allow BuildContext to be used to walk children, just like it can for
  ancestors

- Allow BuildContext to be used to get the Widget of the current
  BuildContext

- Allow StatefulComponentElement to be referenced with a type
  specialisation so that you don't have to cast when you know what the
  type you're dealing with actually is.

Fixes #1084

Now that we've ported all the widget tests to fn3, we don't need a separate fn3
directory.

Also:

 - don't mark a node as from the new generation if it is dirty, since we
   know it still has to be built.

 - establish the rule that you can't call setState() during initState()
   or build().

 - make syncChild() return early for unchanged children.

 - update the tests, including adding a new one.

If it's a StatefulComponent, then it's ok to reuse it so long as it
hasn't been initialised.

If it's a regular Component or a TagNode, then it's always ok to reuse.

If it's a RenderObjectWrapper, then it's ok to reuse so long as it
doesn't have a renderObject.

To put it another way, this changes how we prevent the following
nonsensical situations from arising:

 - Sync two stateful StatefulComponents together
 - Sync two RenderObjectWrappers with RenderObjects together

When either of those cases happen, we just drop the old one on the
ground and use the new one unchanged.