Gpuarray doc

Cuda fftconv

Also add an example of the GPU speedup for borrow to that tutorial.

Make a proxy NVCC_compiler that does not create a second context on the card as a side effect.

Fix: disconnected_inputs arg not getting passed from jacobian to grad

Cuda docs suggest that more than one context on a card may slow down operations, so this might give us a slight win.

This helps immensly when writing or editing documenation by massively
speeding up documentation rebuilds (as in 2-3s versus 60-90s before).

Also sprinkle a bit of pep8.

Add doc for gamma, gammaln, psi and chi2sf.

Fixed error message for scan op dimension mismatch

Implemented basic typed list with some basic op

Added support for images whose last dimension is odd. This is more expensive computationally, and it costs more in memory, but it's optional.

Fixed some typos and a dimension that was probably wrong. This doesn't solve the issue with the boder_mode='full' because we still have other errors.