Expose a random transform's draw with tied sort keys or identity models

To check whether workers repeat a randomized augmentation, make the output reveal the random draw itself. Two tricks: (1) For transforms that sort by a score plus noise (e.g. a randomized Coulomb matrix sorted by row norm plus Gaussian noise), feed copies of a circulant matrix. All row norms tie, so the noise alone picks the permutation, and the flattened output identifies it up to rotations. Then count distinct outputs per in-shard position across shards. (2) For image-augmentation iterators, run the real predict() of a model that is a single Flatten layer, with N copies of one image as input. The predictions are the augmented images, and two match only if the transform parameters matched. Always run a serial or spawn control: it should give about N distinct.