1.fix typo\n2.move import to the top\n3.replace (1-pos_offset_flag) with neg_offset_flag\n4.replace grad.shape[1] with a.shape[1]\n5.remove axis_grad\n Waited to be done rewrite test_extro_ops]

move TypeError into make_node, consequently changed the test a little. Add a note to declarethe difference between fill_diagonal and fill_diagonal_offset

	modified:   ../extra_ops.py
	modified:   test_extra_ops.py
	new file:   test_fill_diagonal_offset.py

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