Related to JDK-8312021, but a point fix for JDK-8252221.

In `MutableSpace::initialize`, we call these two:

```
    if (AlwaysPreTouch) {
      PretouchTask::pretouch("ParallelGC PreTouch head", (char*)head.start(), (char*)head.end(),
                             page_size, pretouch_workers);

      PretouchTask::pretouch("ParallelGC PreTouch tail", (char*)tail.start(), (char*)tail.end(),
                             page_size, pretouch_workers);
    }
```

...which start the pre-touch threads unconditionally, even when we only need one thread anyway. This incurs latency of about 100us (2x 50us) in my setups. We can run the pre-touch the in the caller thread, like it was done before JDK-8252221 to eliminate that overhead.