* typo in c_code (comparing type instead of type.dtype)
* added test case for 3D input tensor in test_ConvOpGrad

Raise a NotImplementedError.

This is because the gradient on the weights is obtained by convoling the visible
layer with the hidden layer (as the kernel). Problem is the convolution code
does not support kernels larger than the input image (which is the case in full
mode). Either fix ConvOp to allow this, or maybe convolve the hidden with the
visible as the kernel, then DimShuffle at the end.

multiple input features

debugmode now checks strides all the time, with the option of raising an exception instead of logging to stderr

- Fixed typo on last line of Python example (y -> u)
- Added note that the first argument is the axis
- Fixed statement saying it is used to concatenate two tensors, while it can
  be used with more
- Fixed some lines too long (>80 characters)