Previously, the batteries-included `apollo-server` package had a special
override of `stop()` which drains the HTTP server before letting the
actual Apollo Server `stop()` machinery begin. This meant that
`apollo-server` followed this nice shutdown lifecycle:
- Stop listening for new connections
- Close all idle connections and start closing connections as they go
  idle
- Wait a grace period for all connections to close and force-close any
  remaining ones
- Transition ApolloServer to the stopping state, where no operations
  will run
- Run stop hooks (eg send final usage report)

This was great... but only `apollo-server` worked this way, because only
`apollo-server` has full knowledge and control over its HTTP server.

This PR adds a server draining step to the ApolloServer lifecycle and
plugin interface, and provides a built-in plugin which drains a Node
`http.Server` using the logic of the first three steps above.
`apollo-server`'s behavior is now just to automatically install the
plugin.

Specifically:
- Add a new 'phase' called `draining` that fits between `started` and
  `stopping`. Like `started`, operations can still execute during
  `draining`. Like `stopping`, any concurrent call to `stop()` will just
  block until the first `stop()` call finishes rather than starting a
  second shutdown process.
- Add a new `drainServer` plugin hook (on the object returned by
  `serverWillStart`). Invoke all `drainServer` hooks in parallel during
  the `draining` phase.
- Make calling `stop()` when `start()` has not yet completed
  successfully into an error. That behavior was previously undefined.
  Note that as of #5639, the automatic `stop()` call from signal
  handlers can't happen before `start()` succeeds.
- Add `ApolloServerPluginDrainHttpServer` to `apollo-server-core`.
  This plugin implements `drainServer` using the `Stopper` class
  that was previously in the `apollo-server` package. The default
  grace period is 10 seconds.
- Clean up integration tests to just use `stop()` with the plugin
  instead of separately stopping the HTTP server. Note that for Fastify
  specifically we also call `app.close` although there is some weirdness
  here around both `app.close` and our Stopper closing the same server.
  A comment describes the weirdness; perhaps Fastify experts can improve
  this later.
- The Hapi web framework has built in logic that is similar to our
  Stopper, so `apollo-server-hapi` exports
  `ApolloServerPluginStopHapiServer` which should be used instead of the
  other plugin with Hapi.
- Remove some examples from READMEs and point to examples in the docs
  instead. Keeping both up to date is extra work.
- Fix some test issues (eg, have FakeTimers only mock out Date.now
  instead of setImmediate, drop an erroneous `const` which made an `app`
  not get cleaned up, etc).

Fixes #5074.