PredictionEngine does not Dispose all resources properly

[tloten]

System Information:

• OS & Version: Windows 11
• ML.NET Version: ML.Net 2.0.1
• .NET Version: .NET 7.0

Describe the bug
PredictionEngine.Dispose() does not dispose of all resources, leaving hundreds of MB of memory behind.
Specifically, I suspect the _outputRow member in PredictionEngineBase is the culprit, as it is not disposed anywhere. Manually disposing of that using reflection hacks causes the expected memory to be free'd up.

To Reproduce
Unfortunately I can't give a code-based repro as I'm using a proprietary ONNX model. However I suspect it's reproducable with other models too.
Steps to reproduce the behavior:

1. Create a pipeline & predictionEngine - I'm using an object-detection ONNX model, being fed a 1280x720 image.
2. Call predictionEngine.Predict()
3. To ensure a clean slate, call GC.Collect() and GC.WaitForPendingFinalizers().
4. Log memory used (560MB in my case)
5. Call predictionEngine.Dispose(), and ensure no references to the PredictionEngine exist any more
6. Call GC.Collect() & log memory (458MB in my case). This still seems very high. Furthermore:
7. Call GC.WaitForPendingFinalizers(); & log memory (72MB in my case). This is now back to where you'd expect it, and suggests that PredictionEngine has left some disposables behind.

Expected behavior
I'd expect the memory usage after calling predictionEngine.Dispose() would drop significantly. I wouldn't expect anything to be left on the Finalizer queue.

Additional context
When I poked around in a memory profiler, it seemed like there were some OnnxRuntime OrtValue items left on the finalizer queue, and they seem to be ultimately owned by the PredictionEngineBase._outputRow field.
So using a reflection hack, I disposed _outputRow myself and it appeared that all memory was being released:

var outputRow = typeof(PredictionEngineBase<TfOdInputData, TfOdOutputData>)
    .GetField("_outputRow", BindingFlags.NonPublic | BindingFlags.Instance)
    .GetValue(predictionEngine);

((IDisposable) outputRow).Dispose();
predictionEngine.Dispose();

Looking at https://github.com/dotnet/machinelearning/blob/main/src/Microsoft.ML.Data/Prediction/PredictionEngine.cs#L129 it seems like disposer should get rid of both inputRow and outputRow.