//
// TF unsorted segment reduction op peforms the following calculation:
//
// Assume segment ids' shape is [SI0, SI1, ..., SIm] and data's  shape is
// [D0, D1, ..., Dn]. Note that segment ids' shape must be a prefix of data's
// shape, so we can have data's shape represented as [SI0, SI1, ..., SIm,
// Dm+1, ..., Dn]. Then
//   output[segment_ids[SI_i0, SI_i1, ..., SI_im], D_im+1, ..., D_in] =