Backchannel doesn't currently guard against Close() being called
multiple times. This can lead to a hang in gRPC when it closes a
broken connection.

Closing a backchannel session shutsdown the backchannel gRPC server.
The gRPC server closes its listener, which is the yamux session.
Closing the yamux session then ultimately closes the underlying TCP
connection. The Close method thus waits for the server to return
the error from Serve to ensure all of the resources have been
cleaned up before returning.

If a connection becomes unhealthy, gRPC will close it from both the
writer goroutine and the keepalive checker. The second one will end
up blocking forever on the error return as the backchannel server has
already been closed and the first closer read the error from the
channel. This commit fixes the issue by running the closing logic only
once and returning the same error on all subsequent calls to close.

Changelog: fixed