fix: restore classical kd signal in variational distillation strategy

Remove incorrect (1.0 - variationalWeight) multiplication that was suppressing
the classical distillation loss and gradient instead of adding variational terms.

This fixes the critical issue where:
- When variationalWeight = 1.0, loss/gradient collapsed to zero
- When variationalWeight < 1.0, only scaled classical KD without variational contribution

Changes:
- ComputeLoss: Remove (1.0 - variationalWeight) scaling from softLoss and combinedLoss
- ComputeGradient: Remove (1.0 - variationalWeight) scaling from soft and hard gradients

Note: This is a baseline fix. Full variational integration (adding variational loss/gradient
weighted by variationalWeight) requires latent representations (mean, logVar) which are not
available in current ComputeLoss/ComputeGradient signatures.

Resolves coderabbitai review comments on lines 63-85 and 87-118.

Author: ooples

src/KnowledgeDistillation/Strategies/VariationalDistillationStrategy.cs

@@ -68,7 +68,7 @@ public override T ComputeLoss(Vector<T> studentOutput, Vector<T> teacherOutput,
         var studentSoft = Softmax(studentOutput, Temperature);
         var teacherSoft = Softmax(teacherOutput, Temperature);
         var softLoss = KLDivergence(teacherSoft, studentSoft);
-        softLoss = NumOps.Multiply(softLoss, NumOps.FromDouble(Temperature * Temperature * (1.0 - _variationalWeight)));
+        softLoss = NumOps.Multiply(softLoss, NumOps.FromDouble(Temperature * Temperature));
 
         if (trueLabels != null)
         {
@@ -78,7 +78,7 @@ public override T ComputeLoss(Vector<T> studentOutput, Vector<T> teacherOutput,
             var combinedLoss = NumOps.Add(
                 NumOps.Multiply(NumOps.FromDouble(Alpha), hardLoss),
                 NumOps.Multiply(NumOps.FromDouble(1.0 - Alpha), softLoss));
-            return NumOps.Multiply(combinedLoss, NumOps.FromDouble(1.0 - _variationalWeight));
+            return combinedLoss;
         }
 
         return softLoss;
@@ -97,7 +97,7 @@ public override Vector<T> ComputeGradient(Vector<T> studentOutput, Vector<T> tea
         for (int i = 0; i < n; i++)
         {
             var diff = NumOps.Subtract(studentSoft[i], teacherSoft[i]);
-            gradient[i] = NumOps.Multiply(diff, NumOps.FromDouble(Temperature * Temperature * (1.0 - _variationalWeight)));
+            gradient[i] = NumOps.Multiply(diff, NumOps.FromDouble(Temperature * Temperature));
         }
 
         if (trueLabels != null)
@@ -109,8 +109,8 @@ public override Vector<T> ComputeGradient(Vector<T> studentOutput, Vector<T> tea
             {
                 var hardGrad = NumOps.Subtract(studentProbs[i], trueLabels[i]);
                 gradient[i] = NumOps.Add(
-                    NumOps.Multiply(NumOps.FromDouble(Alpha * (1.0 - _variationalWeight)), hardGrad),
-                    NumOps.Multiply(NumOps.FromDouble((1.0 - Alpha) * (1.0 - _variationalWeight)), gradient[i]));
+                    NumOps.Multiply(NumOps.FromDouble(Alpha), hardGrad),
+                    NumOps.Multiply(NumOps.FromDouble(1.0 - Alpha), gradient[i]));
             }
         }