Don't use a ThreadGroup for tracking worker threads

[bpot]

Mongrel current uses a ThreadGroup to track its worker threads. But this has the unintended side-effect that any thread spawned by a request will inherit the group of the request thread. Thus, mongrel will think these new threads are requests and take them into account when it enforces num_process restrictions. This mostly went unnoticed for us in 1.1.3 but after upgrading 1.1.5 the num_process enforcement fix caused mongrels to start rejecting requests after we spawned a background thread pool.

I've fixed this by tracking threads in an array and then purging any completed threads whenever this array is accessed. This is not ideal, but I think any other fix would require the thread removing itself from the list of workers and thus require a mutex. Using an array seemed like the most straight-forward fix.