We've had a lot of issues where `stoppable` tests are flaky because they
depend on measuring the time that something takes and hoping it's
relatively close to a particular timeout that's supposed to be
"controlling" the observed behavior.

This PR refactors the internal Stoppable object so that instead of
taking a millisecond timeout value, it takes a DOM-style AbortSignal.
This object is only used by ApolloServerPluginDrainHttpServer; we move
the setTimeout from inside Stoppable into that plugin, which now creates
an AbortController polyfilled from `node-abort-controller` (like we
already are in the usage reporting plugin; note that once we drop Node
v14 support we can switch to the built-in implementation).

Now the Stoppable test can control the grace period directly via
AbortSignals rather than indirectly based on timeouts. This lets us drop
all of the time measurements from the test. There still are some delays
but they are just of the form "wait some time and double-check that
nothing happened": the worst case scenario here is that a test passes
that should fail because the thing would have happened if you'd waited
slightly longer, but there shouldn't be any spurious failures.

Fixes #6963.

….test.ts

Co-authored-by: Trevor Scheer <trevor.scheer@gmail.com>