I wanted to make sure that I get the right timings. It seems that CVM is
faster then what James suggested for now on CPU, and slower on GPU(but can
be fixed). Note that a full implementation of what James suggested will be
slower then what this text predicts ( this is as fast as it will ever get ),
namely scan needs to do a bit more bookeeping that the cython right now
does.

numpy.frompy have changed. I try to don't use it as this should be faster too.

complied with suggestion given by Fred.

This bug was found by the buildbot in debug mode.

They are all working except the outer one because outer doesn't turn into a GEMM
yet.