Add tolerance() functions for MMC classes (#179)

Separate group for the min mean cycle classes (#179)

Simplify comparisons in min mean cycle classes (#179)
using extreme INF values instead of bool flags.

Add HartmannOrlin algorithm class (#179)
This algorithm is an improved version of Karp's original method,
it applies an efficient early termination scheme.
The interface is the same as Karp's and Howard's interface.

Add Karp algorithm class (#179)
based on the MinMeanCycle implementation in SVN -r3436.
The interface is reworked to be the same as Howard's interface.

Rename MinMeanCycle to Howard (#179)

Add a detailed test file for MinMeanCycle and fix test_tools.h (#179)

Rename cyclePath() to cycle() in MinMeanCycle (#179)

Traits class + named parameters for MinMeanCycle (#179)

- Add a Traits class defining LargeValue, Tolerance, Path types.
LargeValue is used for internal computations, it is 'long long'
if the length type is integer, otherwise it is 'double'.
- Add named template parameters for LargeValue and Path types.
- Improve numerical stability: remove divisions from the internal
computations. If the arc lengths are integers, then all used
values are integers (except for the cycleMean() query function,
of course).

Rework and fix the implementation of MinMeanCycle (#179)

- Fix the handling of the cycle means.
- Many implementation improvements:
- More efficient data storage for the strongly connected
components.
- Better handling of BFS queues.
- Merge consecutive BFS searches (perform two BFS searches
instead of three).

This version is about two times faster on average and an order of
magnitude faster if there are a lot of strongly connected components.