fix: normalize logits to probabilities before computing adaptive temperature

In AdaptiveTeacherModel.ComputeAdaptiveTemperature, difficulty was computed
from raw logits which broke bounds. Now:
- Convert logits to probabilities using softmax before difficulty computation
- Clamp resulting difficulty to [0,1] for all strategies
- Ensures temperature stays within [_minTemperature, _maxTemperature] bounds

Author: claude

src/KnowledgeDistillation/Teachers/AdaptiveTeacherModel.cs

@@ -178,18 +178,21 @@ public void UpdateStudentPerformance(int sampleIndex, Vector<T> studentPredictio
     /// </summary>
     private double ComputeAdaptiveTemperature(Vector<T> logits, double baseTemperature)
     {
+        // Convert logits to normalized probabilities
+        var probs = ApplyTemperatureSoftmax(logits, 1.0);
+
         double difficulty = 0;
 
         switch (_strategy)
         {
             case AdaptiveStrategy.ConfidenceBased:
                 // Lower confidence = harder sample = lower temperature
-                difficulty = 1.0 - GetMaxConfidence(logits);
+                difficulty = 1.0 - GetMaxConfidence(probs);
                 break;
 
             case AdaptiveStrategy.EntropyBased:
                 // Higher entropy = harder sample = lower temperature
-                difficulty = ComputeEntropy(logits);
+                difficulty = ComputeEntropy(probs);
                 break;
 
             case AdaptiveStrategy.AccuracyBased:
@@ -198,6 +201,9 @@ private double ComputeAdaptiveTemperature(Vector<T> logits, double baseTemperatu
                 break;
         }
 
+        // Clamp difficulty to [0,1] to ensure temperature stays within bounds
+        difficulty = Math.Max(0.0, Math.Min(1.0, difficulty));
+
         // Map difficulty [0,1] to temperature [min, max]
         // High difficulty (0) -> min temp (sharper)
         // Low difficulty (1) -> max temp (softer)