deps: Tensors 0.50.0 -> 0.50.1 + adapt to Func<Tensor<T>> trace lambdas

Tensors 0.50.1 is the release the open compile-tracer bugs point to;
0.50.0 was the one I initially pinned but it's strictly a prerequisite
step. The API change that matters for us:

  GetOrCompileInference / GetOrCompileTraining accept Func<Tensor<T>>
  (return = plan output tensor) instead of Action (tail-op inferred).

The explicit return fixes the shape-conditional truncation that forced
PR #1167's Predict-bypass workaround: the tracer no longer guesses
which op produced the plan's output, so rank-3 -> rank-4 promotions
and post-forward reshapes no longer get silently dropped.

Three call sites updated to return the last tensor explicitly:
  src/AiModelBuilder.cs:1344           — return nnModel.Predict(input)
  src/Training/CompiledTapeTrainingStep.cs:139  — return computeLoss(...)
  src/Training/CompiledTapeTrainingStep.cs:269  — return computeLoss(...)

Follow-up (separate PR): revert the 11 Predict-override-restore sites
from PR #1167 back to PredictEager now that the tracer actually
captures their forward behaviour correctly.

Build: dotnet build -f net10.0 + net471 both green, 0 errors.

Author: franklinic

Directory.Packages.props

@@ -5,7 +5,7 @@
   <ItemGroup>
     <!-- AiDotNet ecosystem -->
     <PackageVersion Include="AiDotNet" Version="0.113.0" />
-    <PackageVersion Include="AiDotNet.Tensors" Version="0.50.0" />
+    <PackageVersion Include="AiDotNet.Tensors" Version="0.50.1" />
     <PackageVersion Include="AiDotNet.Native.OneDNN" Version="0.50.0" />
     <PackageVersion Include="AiDotNet.Native.OpenBLAS" Version="0.50.0" />
     <PackageVersion Include="AiDotNet.Native.CLBlast" Version="0.50.0" />

src/AiModelBuilder.cs

@@ -1371,9 +1371,11 @@ private void ApplyGradientCheckpointingFromMemoryConfig()
                     try
                     {
                         using var noGrad = new AiDotNet.Tensors.Engines.Autodiff.NoGradScope<T>();
-                        // Discard return: CompiledModelCache treats the last recorded
-                        // op's output tensor as the plan's output.
-                        nnModel.Predict(input);
+                        // Tensors 0.50.1 changed GetOrCompileInference from Action to
+                        // Func<Tensor<T>> — the tracer now binds the plan output to
+                        // whatever the lambda returns, rather than inferring it from
+                        // the last recorded op. Return the Predict result explicitly.
+                        return nnModel.Predict(input);
                     }
                     finally
                     {

src/Training/CompiledTapeTrainingStep.cs

@@ -138,8 +138,12 @@ public static T Step(
                 compositeKey,
                 () =>
                 {
+                    // Tensors 0.50.1 changed GetOrCompileTraining to take a
+                    // Func<Tensor<T>> — the trace lambda must return the
+                    // scalar output (loss) so the compile-graph has a single
+                    // terminal node to differentiate from.
                     var predicted = forward(input);
-                    computeLoss(predicted, target);
+                    return computeLoss(predicted, target);
                 },
                 parameters);
 
@@ -268,8 +272,12 @@ or AiDotNet.Tensors.Engines.Compilation.OptimizerType.Adam
                 compositeKey,
                 () =>
                 {
+                    // Tensors 0.50.1 changed GetOrCompileTraining to take a
+                    // Func<Tensor<T>> — the trace lambda must return the
+                    // scalar output (loss) so the compile-graph has a single
+                    // terminal node to differentiate from.
                     var predicted = forward(input);
-                    computeLoss(predicted, target);
+                    return computeLoss(predicted, target);
                 },
                 parameters);