This is currently a rough sketch and should not be merged. I just wanted to get some feedback here.

The current implementation of oauth refresh offloads the IO to a separate executor, however when a token expires, all calls to getRequestMetadata would hang until the token is refreshed.
This PR is a rough sketch to improve the situation by adding a stale state to token. If a token is within a minute of expiration, the first request to notice this, will spawn a refresh on the executor and immediately return with the existing token. This avoids hourly latency spikes for grpc.

The implementation uses guava's ListenableFutures to manage the async refresh state. Although the apis are marked BetaApi in guava 19, they are GA in guava 30

* The class introduces 3 distinct states:
  * Expired -  the token is not usable
  * Stale - the token is useable, but its time to refresh
  * Fresh - token can be used without any extra effort
* All of the funtionality of getRequestMetadata has been extracted to asyncFetch. The new function will check if the token is unfresh and schedule a refresh using the executor
* asyncFetch uses ListenableFutures to wrap state: if the token is not expired, an immediate future is returned. If the token is expired the future of the refresh task is returned
* A helper refreshAsync 	& finishRefreshAsync are also introduced. They schedule the actual refresh and propagate the side effects
* To handle blocking invocation: the async functionality is re-used but a DirectExecutor is used. All ExecutionErrors are unwrapped. In most cases the stack trace will be preserved because of the DirectExecutor. However if the async & sync methods are interleaved, it's possible that a sync call will await an async refresh task. In this case the callers stacktrace will not be present.

…mplemented on top of 0.8, so now I'm backporting features to minimize the merge conflicts

…and allow subsequent callers to use the stale token

…e caller