Running CircleCI render tests on Windows, Mac and Firefox

[SnailBones]

Closes https://github.com/mapbox/gl-js-team/issues/480

Adds three new runs in CircleCI:

• test-render-macos-chrome
• test-render-windows-chrome
• test-render-linux-firefox

Windows and Mac tests are running on virtual machines.

Failing tests on Windows and Mac are due to bugs including #7331. Using --use-gl-desktop lets these tests pass on Mac but also causes a different set of failing tests due to missing globes, inconsistent stars and fill extrusion patterns (as was the reason for disabling this locally in #12090).

The same tests skipped on Windows and Mac formerly failed on Linux without the --use-gl-desktop flag (added in #10389) as seen in this run without the flag.

I've taken the following approach to updating "allowed" properties:

• If the images are visually indistinguishable and dif is suitably low (< 1% for small images and less for larger or detail-oriented ones), raise the "allowed" threshold just above the dif.
• Otherwise, add to a platform-specific ignore file.
• For cases where appearance is still good but the dif is higher, we may want to consider introducing platform-specific "allowed" values, but I've avoided that for now to minimize the scope of this PR.

Previous runs tended to finish in 5-6 minutes. All of the new runs added take longer:

• Firefox: ~8m
• Mac: ~14m
• Windows: ~20m

To address this, I've implemented parallel runs on Mac and Windows. This has required some restructuring of how reading testfiles work. If a tests-to-run.txt file is present in the root directory, the test runner will read tests from that file instead of running all tests that match a glob.

There's likely room to optimize the CI config through caching. but 90% of the time spent on Windows is running tests. A more effective way to speed up CI outputs would be splitting it into two (or more) runs, with each running half the tests.

Issues tracking new tests in ignores:

• raster-resampling has no effect on raster tiles (eg. nearest neighbor is still blurry) #7331
• render-tests/debug/terrain/collision-occlusion test does not render on Firefox #12406
• Some terrain tests cause time out on Windows CI runs #12407
• render-tests/fit-screen-coordinates/terrain test results in different zoom on CircleCI #12408
• Icon not rendering on Windows and Mac CI runs when icon and screen are different sizes. #12409

Launch Checklist

• Operating systems
  • Windows
  • MacOS
• Browsers:
  • Firefox
• Update (default?) "allowed" values so tests pass
• Open issues tracking all platform-specific test failures
• Performance:
  • [] Cache yarn and other slower setup steps when possible (Caching yarn cache makes runs slower on Windows)
  • Parallelize test runs to bring run time to < 10 minutes

[SnailBones]

This test was incorrectly reverted in #10562 and was passing due to a high "allowed." By removing the allowed and correcting the test we should now correctly catch a regression to order.