fix a bug where we pushdown the max to the input of softmax, softplus, tensor.exp, tensor.log, tensor.tanh, sigmoid, softmax_with_bias.

Add test for this bug and print a warning by default if we encounter the case where the bug was causing problem. The warning can be disabled by the Theano flag warn.argmax_pushdown_bug.

Add test for CAReduce{maximum,mul} and modified CAReduce to handle correctly maximum.

Make the copy to the cpu from the gpu work for empty cuda array.
Make the perform of Sum work in python when their is more then 1 dimenssions