* Fix web test flakiness; enable web golden tests
The fix is three-part:
- Only allow one test to load _and_ test at any point in time.
- Use a fresh Chrome instance for each test file.
- Increase Cirrus resources.
The first two changes only fix the "Unknown error loading" error, but not hanging tests. The resource increase also prevents hanging tests.
Other minor changes:
- Remove test batching (it's no longer necessary)
- Fix the Chrome class, which was using the wrong Completer.
Updates the IOSDevicePortForwarder to no longer depend on context, or on an IOSDevice instance. Instead, it receives all necessary configuration through the constructor.
Moves the IOSDevicePortForwarder to a separate file.
Move Flutter Gallery to dev/integration_tests/ as it is an older copy used only for testing. The current version of the Flutter Gallery now lives in https://github.com/flutter/gallery.
When performing a hot restart, collect isolates without an attached flutter view and send a kill signal. These must have been spawned by running main, so restarting without removing them leads to isolate duplication.
When performing a hot reload, ensure that we send a reloadSources command to every isolate and not just uiIsolates.
Updates VisualStudio and VisualStudioValidator to use constructors instead of global injection. Updates VisualStudio test cases to prefer FakeProcessManager
Updates build_windows test to work without injected VisualStudio
When the AndroidDevice checks for the current API version, it isn't handling a null response. While we can refuse to run on API levels that are unsupported, we should probably assume that a null response = lowest possible API level and allow these to run.
The Azure bot detection can take up to a second to determine if a client is/isn't a bot. To prevent this from slowing down all flutter commands, we can cache the results in the persistent tool state - since we don't expect the same client id to ever become a bot or stop being a bot
Previously the AdbLogReader did async setup in the StreamController.onListen callback, specifically it would query the api version and start the adb process. If the log subscription was cancelled before this setup completed, then the log output could (haven't confirmed) get added to a closed controller, causing the above state error.
