[PyGrain Migration] 1. Output NumPy (not backend tensors) from preprocessing inside Grain pipelines #2945

@preprocessing_function converts eager outputs to backend tensors via convert_preprocessing_outputs.
Reuse Keras core's detection : from keras.src.utils.backend_utils import in_grain_data_pipeline. skip convert_preprocessing_outputs and return NumPy/lists
Ragged results stay lists-of-lists