The problem was that we tested sqr(images2neibs(x)) with too big
values. This caused problem when we compute the numerical gradient.

This will fix some of the stochastic optimization order problem.

Will be used in a later commit.

fixed feature class versus instance issue

Adds correct handling of unimplemented gradients

of instances thereof

disallow uncomputable ops)

added an optimization to stabilize log softmax

optimization

Mixed2

Conflicts:
	theano/gradient.py

Using a Feature to catch BadGradOp

Better doc of extra_ops.

added error checking and doc to reshape