ThreadPoolExecutor statistics methods can be more scalable

We have seen a nice pitfall in some of our services. The highly parallel worker threads reported the metrics for another TPE, asking its getPoolSize(), getActiveCount(), etc. methods. Which lead to a severe contention on `TPE.mainLock`, since all those methods take it:

```
    /**
     * Returns the approximate number of threads that are actively
     * executing tasks.
     *
     * @return the number of threads
     */
    public int getActiveCount() {
        final ReentrantLock mainLock = this.mainLock;
        mainLock.lock();
        try {
            int n = 0;
            for (Worker w : workers)
                if (w.isLocked())
                    ++n;
            return n;
        } finally {
            mainLock.unlock();
        }
    }
```

This contention mode is rather unusual, since one would expect a single monitoring thread to access these methods, but it might not be as easily fixable on application side.

That lock seems to protect the concurrent access to `HashSet workers`. For stable thread pools -- where workers are rarely created, interrupted, or shut down -- we rarely modify it. Maybe we should instead use a ReentrantReadWriteLock here to alleviate this pitfall.