TensorFlowUtils doesn't really fit with the 1.0 API

[rogancarr]

One of the scenarios we want to support in ML.NET 1.0 is the ability to examine the schema of a TensorFlow model (see example below [1]). To do this, we currently use helper methods in the TensorFlowUtils class [2].

Right now, this class and these methods are public and can be used as part of an ML.NET program, but that doesn't seem like the way we want to go in v1.0. What do you all think? Does it make sense to add these methods to mlContext and make them internal/BestFriend? If we add them to a catalog, where should they go?

[1] Example of why you would want to examine the schema before loading. Here, we want to get the layer names and their sizes before specifying them as outputs in our IDataView:

// Get the output size of a layer
var schema = TensorFlowUtils.GetModelSchema(mlContext, model_location);
var outputColumn = schema.GetColumnOrNull("Output");
if (outputColumn.HasValue)
    outputLength = ((VectorType)outputColumn.Value.Type).Dimensions[0];

[2] Helper Methods

public static Schema GetModelSchema(IHostEnvironment env, string modelPath);
public static IEnumerable<(string, string, ColumnType, string[])> GetModelNodes(IHostEnvironment env, string modelPath);
public static TensorFlowModelInfo LoadTensorFlowModel(IHostEnvironment env, string modelPath);

where LoadTensorFlowModel is already in the Transforms catalog.

Would GetModelSchema and GetModelNodes go in the DataOperations catalog?

And to pile on, GetModelNodes (which may not be necessary to expose) returns a list of returns, instead of an object, which we definitely can't have in V1.

Related to #2498

[Ivanidzo4ka]

Related to #2552.
I think main discussion is going on in that issue.