StreamingMR (Title tentative. Suggestions welcome!)

This is a C++ emulated implementation of the planned hardware implementation of
MapReduce.

The key problem we're trying to solve is that MapReduce is difficult to be
effective on the hardware because of the limited RAM capacity for intermediate
storage.

The solution that is proposed is: Instead of storeing K/V pairs in the
intermediate storage, let's store per-key contexts, or local variables, in the
intermediate storage. For classes of applications where the number of keys is
much smaller than the amount of data, this can be effective.
(Histograms might be a good example)

Bear in mind that dynamic memory such as malloc do not exist on the hardware (or
are very slow).

The main work of the programmer is to inherit and implement two classes:
MapReduceApp and InputParser.

This will require inheriting and implementing three data classes:
InputPair, Pair and Context.

InputParser reads a raw data file and turns it into a stream of InputPairs.
MapReduceApp->map is called by the platform for each InputPair, and emits a
stream of Pair's. MapReduceApp->reduce is called with Pair and that Pair's
Context, and modifies the Context state. MapReduceApp->finalize is called with
each Context for all keys once all data is processed, to emit final results.

An example of a random 4 byte word-counting app is provided in
examples/wordcount