This is a significant refactoring of the statistical inference code,
with many cleanups and improvements, and is aimed towards future
extensibility.

The old community module is available as 'community_old' and will be
completely removed in the near future.

This implements a general wrapper for dense_hash_map and dense_hash_set,
that falls back to unordered_map and unordered_set, if sparsehash is not
enabled.

Xin Sun authored Dec 21, 2013 and Tiago Peixoto committed Dec 21, 2013

In case google sparsehash uses a non-standard prefix, e.g., in Homebrew.

Avoid hash of pair<vertex_t, vertex_t> types.

This greatly improves the performance of the previous algorithm, and
includes support for hierarchical models.

Dumping boost::adjacency_list<> improves memory usage by a factor of
two, and also slightly improves performance in some cases.

This enables the inclusion of the non-boost tr1 headers if a recent
enough version of libstdc++ is detected. This als fixes a bug with
clang when sparsehash pulls unwanted tr1 headers.

The condensation_graph() function now returns a property map which
contains the community values themselves for each vertex.

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.

Edges were not properly indexed after graph generation.

Thins changes the graph filtering code slightly to wrap graph types with
GraphWrap, which automatically updates the edge index list when edges
are removed and added to the graph.

This also changes how graphs are passed to algorithms, which is now by
reference instead of pointer. (hence this touches lots of code, but
changes are trivial)

This commit splits libraph_tool into different libraries:
 
   - libgraph_tool_core
   - libgraph_tool_clustering (*)
   - libgraph_tool_community (*)
   - libgraph_tool_correlations (*)
   - libgraph_tool_distance (*)
   - libgraph_tool_generation (*)
   - libgraph_tool_layout (*)
   - libgraph_tool_misc (*)
   - libgraph_tool_stats (*)

It also adds the python sub-module 'test', which provides extensive unit
testing of the core functionality. The core library is fully functional
and all test pass successfully.

(*) -> module needs to be ported to new refactoring, and does not yet build

This is a huge commit which completely refactors the metaprogramming
engine which generates and selects (at run time) the graph view type and
the desired algorithm implementation (template instantiation) that runs
on it.

Things are laid out now as following. There exists a main underlying
graph type (GraphInterface::multigraph_t) and several other template
classes that mask it some way or another, in a hierarchic fashion:

     multigraph_t -> filtered_graph (edges only, vertices only, both)
         |                               |           |           |
         |                               |           |           |
         |-------(reversed_graph)--------|-----------|-----------|
         |                               |           |           |
         \------(UndirectedAdaptor)------------------------------/

The filtered_graph filters out edges and/or vertices from the graph
based on some scalar boolean property. The reversed_graph reversed the
direction of the edges and, finally, the UndirectedAdaptor treats the
original directed graphs as undirected, transversing the in- and
out-edges of each vertex indifferently. Thus, the total number of graph
view types is 12. (The option --disable-graph-filtering can be passed to
the configure script, which will disable graph filtering altogether and
bring the total number down to 3, to reduce compile time and memory
usage)

In general, some specific algorithm, implemented as a template function
object, needs to be instantiated for each of those types. Furthermore,
the algorithm may also depend on other types, such as specific
property_maps. Thus, the following scheme is used:

    struct my_algorithm // algorithm to be implemented
    {
        template <class Graph, class PropertyMap>
        void operator()(Graph *g, PropertyMap p, double& result) const
        {
            // ...
        }
    };

    // in order for the above code to be instantiated at compile time
    // and selected at run time, the run_action template function object
    // is used from a member function of the GraphInterface class:

    double GraphInterface::MyAlgorithm(string prop_name)
    {
        double result;
        boost::any vprop = prop(property, _vertex_index, _properties);
        run_action<>()(*this, bind<void>(my_algorithm(), _1, _2,
                                         var(result)),
                       vertex_scalar_properties())(vprop);
        return result;
    }

The whole code was changed to reflect this scheme, but now things are
more centralized and less ad-hoc code needed to be
written. Unfortunately, due to GCC's high memory usage during template
instantiations, some of the code (namely all the degree correlation
things) had to be split in multiple compilation units... Maybe this will
change in the future if GCC gets optimized.

This commit also touches other parts of code. More specifically, the way
filtering gets done is very different. Now we only filter on boolean
properties, and with the above scheme, the desired implementation runs
with the correct chosen type, and no implicit type conversions should
ever happen, which would have a bad impact on performance.