needed to make this safe for forks from non-main threads

[toregeschliman]

If a process happens to fork from a non-main thread (as happens e.g. in Puma's fork_worker mode) the term_on_timeout counter effectively disappears; whichever thread calls .fork becomes Thread.main in the child process, losing the parent's Thread.main storage. There are a few alternative solutions available: make this a real global (makes people feel icky but that's really what the code is doing) or adding a _fork hook (only practical on recent Rubies or with ActiveSupport required), but the simplest method is to move the ||= 0 initialization code prior to the request-time code that expects it.

Happy to put up a PR with the alternate methods if you'd prefer, please lmk.

(the test_helper change is required for tests to work at all on my local; lots of failures for undefined Rack::Builder otherwise)

[wuputah]

thanks for the contribution!