Fix some GPU restart races.
GpuProcessHost::SendOutstandingReplies will callback the caller with failure,
which may want to retry establishing the channel on a new process right away. So
make sure to be in a consitent state (valid_ = false) in that case.
Also, I observed on Chrome OS at least (haven't been able to repro on Linux
desktop) that the channel failure isn't always detcted right away by the
ChildProcessHost, meaning the request to create a new channel can hang for a
while. So make sure we signal failure as soon as we detect it.
BUG=165849
Review URL: https://chromiumcodereview.appspot.com/
11859006
git-svn-id: svn://svn.chromium.org/chrome/trunk/src@176573 0039d316-1c4b-4281-b951-d872f2087c98