Reduce contention in DFA

[sharwell]

In multithreaded parsing scenarios, contention in DFAState.getTarget and DFAState.setTarget can increase CPU usage dramatically, according to the following report.
http://stackoverflow.com/q/23120555/138304

The following image shows the performance improvement of the rm-contention branch as of commit 57a77e1. The test was performed with 20 threads on a machine with 4 cores + hyperthreading.

[sharwell]

The following shows the performance of commit 5f70f34. The test was performed with 20 threads on a machine with 4 cores + hyperthreading, but the processor is a bit slower than the one from the previous test.

This commit removes synchronization locks from the "fast path" in SparseEdgeMap<T>, and marks some types and methods final to (theoretically) improve the ability of the JIT to inline calls to them.

[sharwell]

Fixed in a86458f