* for APIs like {@link #whenAllComplete whenAllComplete(...)}{@code .}{@link
   * FutureCombiner#call(Callable, Executor) call(...)}, where it is easy to use a new input from
   * the {@code call} implementation but forget to add it to the arguments of {@code
   * whenAllComplete}.
   *

       * regardless of whether some failed. This lets us avoid calling expensive methods like
       * Future.get() when we don't need to (specifically, for whenAllComplete().call*()), and it
       * lets all futures share the same listener.
       *
       * We store `localFuturesOrNull` inside the listener because `this.futures` might be nulled

   * for APIs like {@link #whenAllComplete whenAllComplete(...)}{@code .}{@link
   * FutureCombiner#call(Callable, Executor) call(...)}, where it is easy to use a new input from
   * the {@code call} implementation but forget to add it to the arguments of {@code
   * whenAllComplete}.
   *