Added clamping on conversion functions.

* Shape predictor trainer optimizations

* Fixed performance leak in single thread mode & made VS2010 support

are directly the arguments of a mex function.  This way, we avoid the problem
of dlib::matrix objects being created in other threads that internally call
into matlab and mess it up (since matlab is not thread safe in any sense).

iterators in the sense implied by the C++ standard but rather at least forward
iterators).

* fix shape_predictor dlib::searialize and dlib::desearialize error

* declare shape_predictor searialize and desearialize inline

* fix inline declare issue: declare inline before implement

on windows.