Fix flakiness in HeadlessBrowserTest

In HeadlessContentMainDelegate::RunProcess(), if the headless browser
instance is destroyed before browser runner is shut down, GPU clients
might end up accessing the destroyed object in the meantime. An example
is when a client tries to establish a GPU channel ending up accessing
destroyed HeadlessBrowserImpl::options in order to instantiate
viz::GpuHostImpl. This CL changes the ordering to fix the race
condition.

BUG=881211
TEST=headless_browsertests --gtest_filter=HeadlessBrowserTest.*

Change-Id: I2518eeefcc1e446a59d680fa78eab88e26765404
Reviewed-on: https://chromium-review.googlesource.com/1251331Reviewed-by: Sami Kyöstilä <skyostil@chromium.org>
Commit-Queue: Mohsen Izadi <mohsen@chromium.org>
Cr-Commit-Position: refs/heads/master@{#595465}

headless/lib/headless_content_main_delegate.cc

@@ -282,8 +282,8 @@ int HeadlessContentMainDelegate::RunProcess(
 
   browser_->RunOnStartCallback();
   browser_runner->Run();
-  browser_.reset();
   browser_runner->Shutdown();
+  browser_.reset();
 
   // Return value >=0 here to disable calling content::BrowserMain.
   return 0;