Add graceful shutdown to wasmtime serve, fix flaky CI

[alexcrichton]

This commit fixes the spurious CI failure found here. The problem here is that the infrastructure for testing wasmtime serve was not properly waiting for all output in finish. There's some infrastructure in the tests for spawning a subprocess and managing it, but prior to this commit there was no way to shut down the server gracefully and thus read all pending output from the child tasks.

The specific problem is that a specific error message was expected in the logs after a request had been processed, but the finish method wasn't reading the message. The reason for this is that finish had to resort to kill -9 on the child process as there was no other means of shutting it down. This meant that the print, which happened after request completion, might be killed and never happen.

The solution ended up in this commit is to (a) add a --shutdown-addr CLI flag to wasmtime serve and (b) beef up handling of graceful shutdown. Previously ctrl-c was used to exit the server but it didn't do anything but drop all in-progress work. Instead graceful shutdown is now handled by breaking out of the accept loop and then waiting for all child tasks to complete, meaning that no http requests once received are cancelled. In addition to ctrl-c the --shutdown-addr is used to listen for a TCP connection which is a second means of cancellation. The reason for this is that sending ctrl-c to a process is not nearly as trivial on Windows as it is on Unix, so I didn't want to deal with all the console bits necessary to get that aligned.