The Android webview instrumentation tests seem to be much flakier in multi-process mode.

For example on this bot:
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester?numbuilds=200

https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4026
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4025
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4024
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4023
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4022
https://uberchromegw.corp.google.com/i/chromium.android/builders/Marshmallow%20Tablet%20Tester/builds/4021

Look for tests marked with --webview-sandboxed-renderer:
AwContentsClientFullScreenTest#testOnShowCustomViewAndPlayWithHtmlControl_video with {--webview-sandboxed-renderer}

I discussed this offline with Torne and Toby. They had a few ideas:
- Multi-process may require more resources and on some bots these tests will be more flaky.
- The renderer may be crashing, are we getting mini dumps from it?