Correctly handling variable with batch dimension for math ops.

[emailweixu]

We should allow addition like the following where b is a scalar for each instance in a batch, although both a and b has an additional batch dimension. The following code will cause error with the current implementation of math_op_patch.py.

a = fluid.layers.data(name="a", shape=[2])
b = fluid.layers.data(name="b", shape=[1])
c = a + b

To fix this, we need to set the axis to 0 when the second argument contains batch dimension.

Also makes elementwise ops more tolerant at handling tensors with trailing singular dimensions.
Also added Dim<0>. Dim<0> is for scalar (rank-0 tensor). Adding Dim<0> can simplify a lot of code.

[JiayiFeng]

lgtm