Range overhead reduction.

[akarnokd]

Applied some refactorings and local variable usage to reduce the overhead.

Few observations:

• Having too many local variables may cause register spill, even on x64 which makes some size benchmark faster while other slower.
• The observeOn benchmarks are quite hectic because of receiving thread migration caused by the round-robin worker assignment. It affects the benchmarks with 1 or 1000 elements in the stream.
• Note that the previous OperatorRangePerf size = 1 measured the speed of just due to the optimization of range(). The updated perf now instantiates the OnSubscribeRange.
• Note that the observeOn benchmark with size = 1 run the just() as well.

Benchmark comparison (i7 4770K, Windows 7 x64, Java 8u51)

[akarnokd]

Updated the PR to fix the lack of widening in the fast path if the end is Integer.MAX_VALUE by adding 1L (the slow path adds the long idx value which does this widening).

[akarnokd]

I'm merging this; the values are better for the thread-stable range perf and observeOn is simply too hectic with the round-robin core usage.