Support for DepthToSpace operator

[etiotto]

DepthToSpace rearranges data from depth into blocks of spatial data. This is the reverse transformation of SpaceToDepth. The output tensor is a copy of the input tensor where blocks of size blocksize * blocksize from the depth dimension are moved in spatial blocks to the height and width dimensions. By default, mode = DCR. In the DCR mode, elements along the depth dimension from the input tensor are rearranged in the following order: depth, column, and then row.

Given input tensor [N,C,H,W] by default (mode = DCR] the output tensor has shape [N, C / (blocksize * blocksize), H*blocksize, W*blocksize]. The transformation can be seen as a sequence of a reshape + transpose + reshape operations:

  tmp1 = reshape(input, [N, blocksize, blocksize, C // (blocksize**2), H, W])
  tmp2 = transpose(tmp1, [0, 3, 4, 1, 5, 2])
  out = reshape(tmp2, [N, C / (blocksize**2), H * blocksize, W * blocksize])

When mode = CRD:

  tmp1 = reshape(input, [N, C / (blocksize ** 2), blocksize, blocksize, H, W])
  tmp2 = transpose(tmp1, [0, 1, 4, 2, 5, 3])
  out = reshape(tmp2, [N, C / (blocksize ** 2), H * blocksize, W * blocksize])