Run several independent predictions in one call.

Each item is scored with the same logic as POST /predict (its own model + input).
Items do not pass values to each other. Each successful item writes its own
activity row, same as a single prediction. One item failing does not abort the rest.