Make AMP tests less flaky

[rsimha]

The number of test failures / flakes during PR and push builds on Travis due to dirty global state appears to have increased over the recent weeks. It's time to explore ideas that will reduce flakiness by isolating tests from one another and eliminating dirty global state.

Some ideas:

• Update beforeEach and afterEach in test/_init_tests.js to be more thorough with how they clear state between tests
• Prevent tests from performing operations that modify global state
• Fail tests that when errors are logged by the runtime (unless they're expected) (🏗 Surface console errors during tests as mocha failures #14336)
• Create a new window object (with a new this) for each test (inefficient? infeasible?)

(More ideas welcome via comments below)

[alanorozco]

1. Stub experiment toggles by default

Some sort of test-time replacement of isExperimentOn and toggleExperiment so that they use stubs instead of affecting the window state. Tests that need to validate the behavior of these functions can be whitelisted.

2. Forbid experiment toggles in test

Disadvantage: tests need to be rewritten to stub a static method that checks for the experiment.

3. Force all services to be stubbed in tests

Services are likely to affect global state since they're singletons. Implement some mechanism (similar to above) to either:

1. automatically mock services or
2. forbid the direct use of services without mocks

This would also reduce the running time of a lot of tests.

[dreamofabear]

AKA return of the prodigal son: go/amp-test-excellence

/cc @lannka 😂

[rsimha]

It turns out that several AMP tests do asynchronous stuff (like throwing errors), but don't correctly use the done() callback to indicate that the asynchronous tasks have completed.

This is being tracked by #15658

[dreamofabear]

FR: Run all tests in the scope of a child iframe such that window, document and other globals don't reference the test runner's window.

Motivation: #16586