More options for run() in scaling MCF algorithms (#180)

- Three methods can be selected and the scaling factor can be
given for CostScaling.
- The scaling factor can be given for CapacityScaling.

Entirely rework CostScaling (#180)

- Use the new interface similarly to NetworkSimplex.
- Rework the implementation using an efficient internal structure
for handling the residual network. This improvement made the
code much faster.
- Handle GEQ supply type (LEQ is not supported).
- Handle infinite upper bounds.
- Handle negative costs (for arcs of finite upper bound).
- Traits class + named parameter for the LargeCost type used in
internal computations.
- Extend the documentation.

Port CostScaling from SVN -r3524 (#180)

Traits class + a named parameter for CapacityScaling (#180)
to specify the heap used in internal Dijkstra computations.

Entirely rework CapacityScaling (#180)

- Use the new interface similarly to NetworkSimplex.
- Rework the implementation using an efficient internal structure
for handling the residual network. This improvement made the
code much faster (up to 2-5 times faster on large graphs).
- Handle GEQ supply type (LEQ is not supported).
- Handle negative costs for arcs of finite capacity.
(Note that this algorithm cannot handle arcs of negative cost
and infinite upper bound, thus it returns UNBOUNDED if such
an arc exists.)
- Extend the documentation.

Port CapacityScaling from SVN -r3524 (#180)

Memory leak bugfix in BellmanFord (#51)