This fixes compilation with older GCC versions (ca. 5.0.0).

This changes the underlying adjacency matrix implementation to a single
mixed in + out edge list for each node, instead of separate ones. This
simplifies the undirected_adaptor<> so that its out-edge iterators
become trivial, and also faster.

Now algorithms rely more on parallel_{vertex/edge}_loop() and global
the paralellization threshold is honored.

This also implements direct manipulation of OpenMP settings (schedule
and number of threads) via python.

Use schedule(static, 100), to reduce overhead with small graphs in the
majority of cases.

This enables support for older versions of GCC (>=4.2), by falling back
to the boost tr1 implementation when the GCC version is less than 4.4.

absolute_trust() is now called trust_transitivity(), and the whole
trust matrix can now be efficiently calculated.