(speed tests still to be done)

The current implementation is still somewhat slow because it is not fully in C.
However, I have now implemented a version of the munkres algorithm as a C function,
which should speed up UTED once it is properly integrated.

revised documentation and copyright header in all files to be consistent with the numpy documentation convention

extended the setup file to compile c files if cython is not available and added a MANIFEST file to ensure inclusion of *.pyx files in source distributions

made BEDL a little easier to use by providing a _delta property with the learned delta function and improving the performance of the delta function.

implemented multiprocessing version of backtracing and improved multiprocess handling by also re-raising exceptions in worker threads.