Update perf test shutdown to wait for all operations.

When tearing down the Dawn perf tests we're seeing some racing which
appears to terminate the tests on the bots, there is no logging output
when this happens, just the test fails to teardown.

It seems that there maybe a race between GPU commands and test tear
down.

Specifically:

1. UniformBufferUpdatePref tests do MapAsync on a bunch of buffers, the
callback for those map asyncs uses the `this` pointer to set some
information
2. The test "completes" while the MapAsync messages are in progress
3. The test harness waits for its iteration counter, based on queue
submission. That iteration counter can complete before the map async.
4. The test harness exits the test is torn down and the fixture for the
`this` pointer is destroyed. 5. the MapAsync finishes, tries to access
this and then dies.

This is a hypothesis, I haven't been able to reproduce the test locally
where it crashes.

This CL changes the waiting for a specific iteration count to a
`WaitForAllOperations` call instead. This seems, when trying to run the
tests on the perf bot in the try jobs, to make the flake go away over a
number of runs.

Change-Id: I808a915973eabeb3568b4490ce808c5649543876
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/295355
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Kai Ninomiya <kainino@chromium.org>