This ensures that the remote listener has access to any platforms that
are dynamically loaded in the test runner, so they can be used in
platform selectors.

See #99
See #391

We want users to be able to dynamically define new platforms, which
means we need infrastructure in place for piping those platforms to
places that previously assumed TestPlatform.all was a full list of
available platforms. PlatformSelector is the trickiest example, since
it's parsed in a number of different places and needs to provide useful
feedback to users when they use an undefined platform.

This splits parsing and platform validation into two separate steps.
Validation will be done immediately after parsing when the selectors
come from top-level annotations or parameters passed to test() or
group(), but selectors defined in configuration files are now parsed
only after all configuration is parsed. This will allow new platforms to
be defined *and* referenced in configuration files.

See #99
See #391

This will make it easier to define custom test platforms (both via #391
and #99) in the future. Because those platforms will be loaded based on
the configuration, requiring knowledge of them to parse the user's
platform arguments would produce an unresolvable circular dependency.

This makes it possible to commit reviewed code to a branch when working
on a larger change that's composed of multiple pull requests.

Closes #688

* Adapt for strong-mode changes to zone.

* Use `runGuarded` instead of `runUnaryGuarded`.

* Increment version number.

* Add Changelog entry.

* Drain and ignore

* add changelog and pubspec

Fixes #667

Fixes #674

dynamic is treated as the bottom type for function arguments, but this
will soon change. Null will continue to be the bottom type.

The question we are asking is "Is this a function that takes one
argument", when dynamic is no longer the bottom type the question will
become "Is this a function that takes one dynamic argument", and a
function which take one argument of a specific type does *not* take a
dynamic argument.

* extend timeout

Bump to 0.12.24+1.

This also makes CompilerPool more general, so it can be used by both
the node and the browser platforms.

This is redundant with StreamQueue.startTransaction().

Run tear-downs in the same error zone as tests.

We used to create a new error zone for each call to
waitForOutstandingCallbacks(), which meant that asynchronous errors
couldn't be safely passed around within a single test. Now we have a
single error handler that manually checks the zone in which the error
was thrown.

Enable associated lint
Verified the new lower bound (v0.26.4 from Dec 7, 2015) works