This change reduces the overhead of calling Ruby.getCurrentContext
since we still call it in a number of places.

* ThreadService now extends ThreadLocal rather than aggregating
  a ThreadLocal in a field. This eliminates one hop.
* All hot-path methods in ThreadService are static.
* Restore @kares recursion logic since it does appear faster than
  an explicit null check loop.

On OpenJDK 8 C2 this reduces single-thread getCurrentContext time
from around 4ns to around 3.2ns. Other VMs have similar gains.

Tested with a trivial benchmark:

```java
import org.jruby.Ruby;

public class ContextGetter {
  public static void main(String[] args) {
    Ruby runtime = Ruby.newInstance();
    while (true) {
      long nanos = System.nanoTime();
      for (int i = 0; i < 100_000_000; i++) {
        runtime.getCurrentContext();
      }
      System.out.println((System.nanoTime() - nanos) / 100_000_000.0);
    }
  }
}
```