Highlight selfGradient

Also remove redundant copy() call.
`Tensor:scalarMultiply` is copying internally.

Author: FlorentinD

ml/ml-core/src/main/java/org/neo4j/gds/ml/core/functions/EWiseAddMatrixScalar.java

@@ -52,10 +52,11 @@ public Matrix apply(ComputationContext ctx) {
 
     @Override
     public Tensor<?> gradient(Variable<?> parent, ComputationContext ctx) {
+        Matrix selfGradient = ctx.gradient(this);
         if (parent == matrixVariable) {
-            return ctx.gradient(this);
+            return selfGradient;
         } else {
-            return new Scalar(ctx.gradient(this).aggregateSum());
+            return new Scalar(selfGradient.aggregateSum());
         }
     }
 }

ml/ml-core/src/main/java/org/neo4j/gds/ml/core/functions/ElementSum.java

@@ -45,6 +45,7 @@ public Scalar apply(ComputationContext ctx) {
 
     @Override
     public Tensor<?> gradient(Variable<?> parent, ComputationContext ctx) {
-        return ctx.data(parent).map(ignore -> ctx.gradient(this).value());
+        double selfGradient = ctx.gradient(this).value();
+        return ctx.data(parent).map(ignore -> selfGradient);
     }
 }

ml/ml-core/src/main/java/org/neo4j/gds/ml/core/functions/L2NormSquared.java

@@ -54,6 +54,7 @@ public Scalar apply(ComputationContext ctx) {
 
     @Override
     public Matrix gradientForParent(ComputationContext ctx) {
-        return ctx.data(parent).copy().scalarMultiply(2 * ctx.gradient(this).value());
+        double selfGradient = ctx.gradient(this).value();
+        return ctx.data(parent).scalarMultiply(2 * selfGradient);
     }
 }

ml/ml-core/src/main/java/org/neo4j/gds/ml/core/functions/ReducedCrossEntropyLoss.java

@@ -69,6 +69,7 @@ public static long sizeInBytes() {
 
     @Override
     public Scalar apply(ComputationContext ctx) {
+        // manually call forward as `predictions` is not registered as a parent
         var predictionsMatrix = ctx.forward(predictions);
         var labelsVector = ctx.data(labels);
 
@@ -85,6 +86,7 @@ public Scalar apply(ComputationContext ctx) {
 
     @Override
     public Tensor<?> gradient(Variable<?> parent, ComputationContext ctx) {
+        // manually call forward as `predictions` is not registered as a parent
         var predMatrix = ctx.forward(predictions);
         var labelsVector = ctx.data(labels);
         int numberOfExamples = labelsVector.length();