Fix memory leak in TensorflowTransform

[KsenijaS]

This PR fixes two memory leaks:

1. After executing the graph, input tensor are not being freed
2. When running multiple pipelines (sessions), graphs are not being freed.

This fixes #4134

[dnfclas]

All CLA requirements met.

[eerhardt]

I think this approach is fine for now. But as I discussed with @codemzs, we need to solve this issue all up in ML.NET. See #906 (which was closed for some reason?).

TensorFlow and Onnx transformers both have native resources that need to be cleaned up. And with #4157 we are adding even more transformers that require native resource clean up. Relying on Finalizers to do this cleanup is not ideal. Instead, we should allow users to dispose transformers that need disposing without relying on the Finalizer to do it.

[codemzs]