This is a large commit which replaces lambda::bind with boost::bind in
most parts of the code. This improves compilation time, and slightly
decreases compilation memory usage in some cases.

This includes the edmonds karp, push relabel and kolmogorov maximum flow
algorithms from the BGL.

Use correct types and always pass GraphWrap objects by value.

This also includes convenience property map checks.

load() function no longer supported in scipy.

No longer only thrown GraphError upon any error, but instead throw
specific exceptions which are more meaninful and are mapped to standard
python exceptions, such as IOError, ValueError and RuntimeError.

Path sampling is done using different bias, and the edge sampling is now
done correctly. There is now only one parameter 'n_iter' which replaces
'min/max_iter' and 'epslon'.

When using undirected graphs, the implicit direction of each edge was
not properly handled, and the correlated case would not work. Now, in
this case, the direction is always checked, and the edge is inverted if
necessary.

This would cause the graph to become directed (when previously
undirected) in some circumstances.

This is a workaround for a behaviour of graphviz which is not well
understood.

The RNG object must be passed by _reference_ not by value, as is the
default! Otherwise the RNG state does not change...

Target degrees were not being sampled correctly from their probability.

Reference counts of newly created arrays must always be incremented, and
not borrowed().

This prints useful library information.

We will rely on sphinx + docstrings instead.

Import numpy and scipy before everything to avoid weird segmentation
faults depending on the order things are imported.