[C++] Potential race condition in readahead generator

It is possible for the final future to get marked finished twice.

Thread 1: Polls the generator, checks finished (false), is about to add a task
Thread 2: Finishes, marks finished, decrements running tasks to 0
Thread 1: adds task, immediately finishes, decrements running tasks to 0

Both threads will attempt to mark the final future finished.

**Reporter**: [Weston Pace](https://issues.apache.org/jira/browse/ARROW-17996) / @westonpace

_{**Note**: *This issue was originally created as [ARROW-17996](https://issues.apache.org/jira/browse/ARROW-17996). Please see the [migration documentation](https://github.com/apache/arrow/issues/14542) for further details.*}

No file or test is named. Locate the C++ readahead generator and inspect the final-future completion path, then reason through or reproduce the two-thread interleaving described. Done means concurrent completion cannot mark the final future finished twice, with a regression test covering the race.

Written by the indexing model from the issue text.