Create tests to prove recover of gRPC keep-alive

[pingles]

We've noted in other issues (#425 and #217) that the gRPC load-balancing behaviour can get itself into a state where it won't update server addresses.

The gRPC team's advice in other threads (#425 (comment)) is to use the server keep alive to force the client to disconnect and refresh addresses periodically.

Although we do have some tests that stand up a gRPC server and client, it'd be good to simulate this failure mode and prove the handling via the gRPC config settings.

I started the work in #441 but haven't had time to progress.