Cleanup RunningAverage implementations

[bertm]

the RunningAverage implementations have various issues:

• Too many responsibilities (like TimeDecayingRunningAverage, that tries to keep track of clock skew in the middle of its calculations)
• Too much mutable state (like SimpleRunningAverage, that keeps track of "how many items do I have" across 3 different fields)
• A generous amount of synchronization, but not everywhere
• Leftover methods from earlier times (custom serialization formats and unused code that does who-knows-what)
• A large amount of empty Javadoc comment
• A general lack of consistency in code formatting
• Zero tests

This PR attempts to improve the situation by:

• Separating out the clock skew tracking & reporting out of the main logic of TimeDecayingRunningAverage (and using a monotonic clock for the actual computations)
• Keeping the state in a single immutable object accessed through an AtomicReference where possible, so state is in one place and can easily be reasoned about
• Replacing synchronization with compare-and-set operations on the reference
• Removing the legacy format read and write method, and removing Serializable from the class hierarchy (this is technically a breaking, but in my view it is highly unlikely that anybody would be using these methods)
• Removing the empty Javadoc. Most of the operations are explained on the RunningAverage interface; others are sufficiently clear.
• Reformat everything (consistent use of tabs vs spaces, yay!), move constructors to the top, etc.
• Add some tests to cover most of the functionality

Also, deprecate MedianMeanRunningAverage as it's probably never a good idea to use it.