Fix compatibility with parallel stream (master)

[szpak]

identity value provided in reduce is reused across threads and the accumulation operation needs to work on a copy. For more memory efficient reduce operation a dedicated Collector with mutable internal accumulator can be used.

Also added tests to verify that custom user implementation is parallel stream compatible and a logger to show in practice how values change in each step.

[nurkiewicz]

As a second thought (also to #5), do we really want to even talk about parallel streams here? They are mentioned briefly several lessons later, only to tell they are a bad idea. I'm reluctant to test them here. Also the test code is much more complex then before, which makes learning curve steeper.