The Iris classification sample way of reusing the "Label" column is confusing

As mentioned by this ML.NET user, the way the [Iris Multiclassification sample](https://github.com/dotnet/machinelearning-samples/tree/master/samples/csharp/getting-started/MulticlassClassification_Iris) reuses the "Label" column is confusing:

https://github.com/dotnet/machinelearning/issues/5264#issuecomment-652487403.

The problem is that in the sample, we have this:
```C#
.Append(mlContext.Transforms.Conversion.MapKeyToValue(outputColumnName: nameof(IrisData.Label) , inputColumnName: "KeyColumn"));

```
Which the user interpreted that it was mapping the result of the prediction to the `Label` field.

What that step is doing is to actually map the key version of `Label` back to the original value of `Label` itself. Still, I don't know why this step is necessary, since the Label column itself is never modified (the MapValueToKey used at the beginning of the pipeline is actually mapping the Label column into a new KeyColumn, so no need to map back those values).

The lack of a `MapKeyToValue` that actually mapped back the values of "PredictedLabel" was what confused the user.

Furthermore, I also find strange that later when [consuming the model](https://github.com/dotnet/machinelearning-samples/tree/master/samples/csharp/getting-started/MulticlassClassification_Iris#4-consume-model) the approach to get the values for the key "PredictedLabel" column is to actually get the annotations. Which is somewhat a weird approach, since I'd find using a `MapKeyToValue` to be more user-friendly.

I guess a closer look to this is needed to determine if there's a reason behind those decisions. But at first look it seems that they're unnecessary, and it would be less confusing for users if we get rid of them. Thanks 😄