This removes the need to manually include the dialog builder in the main window's build() function.
It also removes the need to track if a dialog is visible.

Other changes:
- I made dialog.dart a bit more readable.
- I renamed transitionFinished to fullyOpaque since that's what actually matters.
- I made Routes track if they're opaque. Eventually this should probably be more configurable when creating the route.

Directions for Future Research:
- Use this for focus management somehow.
- The popup menu should use something like this.
- We should factor the following out into a showDialog() function that returns a future for the dialog's exit result:
    navigator.push(new DialogRoute(builder: (navigator, route) { ... }));
- Maybe navigator.pop() should take a value to return to that Future.

Converted demo launcher example to library

Converted fitness example to library

Converted hello world example to library

Converted mine digger example to library

Converted stocks example to library

Roll was ff8bbe4e
This is another step in fixing #138

R=viktork@google.com

The workbench package exists to pull in pub packages from both pub.dartlang.org
and from the source tree (using dependency overrides). The idea is that
workbench will reflect a typical Sky developer's environment so that we can use
the same tools as a typical sky developer by running them with the workbench as
the cwd.

Turns out if we don't have these deps, nothing builds these packages.

We now just symlink the lib directory into the packages directory in the output
directory.

This package contains the mojom.dart files we generate from the sky_engine
repository.

After this patch, we use pub to manage versioning and fetching the material
design icons.

This CL introduces a SkyShell.apk that is separate from the SkyDemo.apk that we
upload to the store to show our demos.  The SkyShell.apk is just an empty shell
that can run Sky applications on Android.

I've noticed an anti-pattern emerge where people call scheduleBuild()
when they think they've changed enough state that they should rebuild,
instead of just wrapping their changes in setState(). This leads to
them missing state changes, having extraneous scheduleBuild() calls,
and other similar bugs.

By removing scheduleBuild(), the only way to actually schedule a build
now is to call setState(), and hopefully that'll make it much clearer
that you're only supposed to do this when you change state.

Assert that there are no duplicates.
Export GlobalKey from basic.dart, so that people don't have to import widgets.dart just for that.
Fix the "initialFocus" feature which actually didn't work.

This is supposed to make Viktor's game faster, but it's not clear
to me that it actually does.  I've left the code
using the dart version of drawAtlas for now until Viktor can
verify that it looks correct.

I also added a wrapper for SkFilterQuality in the process of
debugging SkCanvas.drawAtlas since all drawAtlas examples
in Skia use FilterQuality.low.  The bug which blocked me for
so long turned out to be that SkCanvas.drawAtlas doesn't
draw anything if antialiasing is turned on.

Issue #138.

R=abarth@google.com

RenderConstrainedBox was ignoring its additional constraints for its intrinsic size logic when it had no child.

 - Remove unused imports
 - Remove unused, broken constructor for MineDiggerApp

This patch provides a class to manage focus, Focus, and a class to
manage ownership of the keyboard, KeyboardHandle.

Inherited (in widgets.dart) is adjusted to support subclasses that
need to copy state from instance to instance.

A GlobalKey key type is introduced that is basically the same as
UniqueKey. Component classes that need a globally-unique key can
specify that their 'key' constructor argument is a GlobalKey.

Focus
-----

You can use `Focus.at(this)` to determine if you, a Component, are
currently focused.

You can use `Focus.moveTo(this)` to take the focus or give it to a
particular component.

For this to work, there has to be a Focus class in the widget
hierarchy.

Currently, there can only be one Focus class, because nested scopes
aren't supported. We should add support for that in a future patch.
See issue #229.

KeyboardHandle
--------------

Instead of directly interacting with the keyboard service, you now ask
for a KeyboardHandle using `_keyboard.show(client)`. This returns a
KeyboardHandle class. On that class, you can call `handle.release()`
when you want to hide the keyboard. If `handle.attached` is still
true, and you need to reshow the keyboard after the user hid it, then
you can can `handle.showByRequest()`.

The point of this is that the `keyboard.show()` method maintains the
invariant that only one KeyboardHandle is live at a time.

There are some issues with the keyboard service that I filed as a
result of doing this patch: #226 #227

Accessing the network created a race condition that would sometimes cause other
tests to fail.

The Python HTTP server was caching too agressively. Instead, use an HTTP server
written in Dart from the sky_tools package.

If your constraints are tight when you get laid out, you don't get a
relayout subtree root.

If you don't have a relayout subtree root, and you get marked dirty,
you go through layoutWithoutResize() rather than layout(), so we don't
get a parentUsesSize.

If you're not dirty and your constraints didn't change, layout() skips
your layout.

So then if your initial layout had parentUsesSize:true, and then you
got marked dirty directly, you would set your size with
parentCanUseSize=false, and then later if your parent tried to lay you
out then read your size, it would crash because your size wasn't set
up to allow you to get your size.

The fix is to actually remember the last setting of parentUsesSize,
even in the case of the constraints being tight and you later being
marked as needing layout directly.

In theory, before, if you had the same image twice in a scrolling container, you'd get an assertion with no way around it.
This makes those nodes not bother making keys by default, which is cheaper and more correct.

This fixes some theoretical bugs whereby we were using hashCode to try
to get unique keys for objects, but really we wanted object identity.
It also lays the groundwork for a new GlobalKey concept.

I tried to keep the impact on the code minimal, which is why the "Key"
constructor is actually a factory that returns a StringKey. The code
has this class hierarchy:

```
   KeyBase
    |
   Key--------------+---------------+
    |               |               |
   StringKey    ObjectKey       UniqueKey
```

...where the constructors are Key and Key.stringify (StringKey),
Key.fromObjectIdentity (ObjectKey), and Key.unique (UniqueKey).

We could instead of factory methods use regular constructors with the
following hierarchy:

```
   KeyBase
    |
   LocalKey---------+---------------+
    |               |               |
   Key      ObjectIdentityKey   UniqueKey
```

...with constructors Key, Key.stringify, ObjectIdentityKey, and
UniqueKey, but I felt that that was maybe a more confusing hierarchy.
I don't have a strong opinion on this.