Disallow creating bad classification iterator (#8042)

When doing classification we need to know the `numPossibleLabels`. If it's set to -1, then we get obscure and confusing null-pointers when accessing labels when calling `ComputationGraph.fit` on the iterator. This PR blocks the user from shooting themselves in the foot.

deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/src/main/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetIterator.java

@@ -160,6 +160,11 @@ public class RecordReaderDataSetIterator implements DataSetIterator {
      */
     public RecordReaderDataSetIterator(RecordReader recordReader, int batchSize, int labelIndexFrom, int labelIndexTo,
                                        boolean regression) {
+        if (!regression) { 
+            throw new IllegalArgumentException("This constructor is only for creating regression iterators. " +
+                                               "If you're doing classification you need to use another constructor that " + 
+                                               "(implicitly) specifies numPossibleLabels");
+        }
         this(recordReader, new SelfWritableConverter(), batchSize, labelIndexFrom, labelIndexTo, -1, -1, regression);
     }