We were using Dartium settings due to what I assume was a copy-paste
error on my part. I don't think this actually affected anything but
tests, since when running tests we never used the default
ExecutableSettings.

This will allow us to run those tests with Dart 2, where Dartium isn't
supported.

The memory leak occurs when window.parent is accessed through
dart:html, so we access it directly through JS interop instead.

Closes #724

We weren't destroying our references to `iframe`s after removing those
frames from the DOM, so we would slowly accumulate more and more
frames as time went on.

Partially addresses #724

Closes #753

Prepare for fix of https://github.com/dart-lang/sdk/issues/31963

Closes #742

This folds frames from core libraries and packages we don't care
about, but leaves test frames as-is.

Also fix a related documentation error.

Closes #740

Bump to 0.12.29+1 before publishing.

Fix a couple of strong mode runtime cast errors.

In strong mode, a generic type instantiated with dynamic is not a
subtype of all types. You can't pass a List<dynamic> to something
expecting, say, List<int>.

These errors are usually detected statically, and most of those have
been fixed. However, sometimes this becomes a runtime cast, as in:

main() {
  // Store a List<dynamic> in a variable of type dynamic.
  dynamic d = [];

  // Implicit runtime downcast from dynamic to List<String>.
  List<String> s = d;
}

In order to ease the migration to strong mode, DDC has been ignoring
these cast failures when they involve certain commonly used types. We
are now in the process of actively fixing those errors.

More context: https://github.com/dart-lang/sdk/issues/27223

Now enabled in analyzer as of
https://github.com/dart-lang/sdk/commit/0a6253312f
Dart SDK 2.0.0-dev.12.0

We were previously communicating over standard in/out, but this was
fragile: if raw JS code printed anything to stdout, it would break our
connection.

Closes #726

Add flushEventQueue() and neverCalled

Closes #417

We were just using const Object() as a default argument value, but
it's plausible that a user might naturally pass that in to a callback.
Now we use a class that's private to test so that external users will
never use it.

Fix load suite timeout weirdness

Closes #390

This can never fire before the load suite timeout. Now that that
timeout actually works, it's redundant.

See #390

* Fix fuzzy arrow warnings since they will become errors before too long.

* Keep type on _onMessage() parameter.

We weren't completing the _suiteAndZone completer, which meant that
errors (such as timeouts) were reported but didn't actually stop the
suite from loading.

See #390

Make addTearDown() play nice with setUpAll()

addTearDown() now adds a tearDownAll() rather than adding a plain
tearDown() when called from within setUpAll() (or tearDownAll()). This
is technically a behavioral change, but the old behavior was so
useless I doubt anyone was relying on it.

Closes #712

This allows us to avoid issues like those described in #713 where we
want users to be able to share errors between different logical tests.

Use Stdout.terminalColumns for line length

Fall back on a 200-character default, which is less likely to obscure
test names when printing to a non-terminal destination.

Closes #86