small fix: load at the module level, not in the sub function.

I merge it myself as this is a small fix that break the cuda code. Don't hesitate to review it.

make compatible with python 2.4 and check that it was not compiled while...

Updated the vision.

Clone get equiv

Add the Theano op Bartlett, FillDiagonal and Fourier.

Added GammaLn and Psi functions

fix bugs for arctanh

Improve gemm optimizer to never accidentally duplicate work.

Fix scan pushout

When running on GPU, and constructing the list of inputs, they might not end
up on the right device (i.e. the input is on gpu instead of cpu or the other
way around). The simple fix is just to filter the arguments.

Don't hide the argument to Fourrier in fft() and simplify the infer_shape graph when the axis is a constant.