When a gradient is zero because the variable has no influence on the result, now return a variable with same shape instead of a 0 scalar, so that we can rely on the shape being unchanged

added a disabled version of Mean op. That is to make the graph of compiled fct easier to read instead of having multiple op for this fct. Need to make the gradient work before enabling by default.

moved the test_conv.py:TestConv2D test from fct test_basic to fct test_img_kernel_same_shape. This way it don't show up as the basic test that fail. The cause is the gcc on fc9 fail to properly compile that test when the image and kernel have the same size.

Modified theano._asarray to handle int64 as well, and throw an exception for types other than int32 and int64 if a type mismatch is detected. This fixes test theano/sparse/tests/test_basic.py:test_structureddot.test_opt_unpack

Implemented gradient of CrossentropySoftmaxArgmax1HotWithBias when there is an incoming gradient on the softmax's output