Add support for tensors with 0 in shape

[dsmilkov]

TF Python supports tensors that have 0 in their shape (e.g. a 3D tensor with shape [1,0,5]).

Tensors with 0 in their shape have no elements, but broadcasting and concating should still work. E.g. concating [0,N] with [5,N] should return [5,N], and summing [0,N] with [1,N] should give [0,N] (broadcasting 0 with 1 results in 0).

Why have 0 in shapes?
The need arises in real models like ssd_mobilenet_v2_coco where only bounding boxes with confidence above some threshold are retained (see here). Thus, if no boxes have confidence above a threshold, the resulting tensor should be of shape [0, 4] (4 comes from the 4 corners of the bounding box).