Limited functionality for fine-tuning TF models using the TensorFlowTransform

The TensorFlowTransform in ML.NET provides some basic capabilities for re-training pre-trained TF models.

The table below lists down the capabilities currently supported for fine-tuning TF models. It also lists functionality that's currently not supported.

| Fine-Tuning abilities | Supported by ML.NET TensorFlowTransform |
| ---------- | -------- |
| Continue retraining on TF SavedModel, given dataset with same shape as original model | Yes |
| Continue retraining on TF SavedModel, given dataset with _different_ shape as original model | No |
| Freeze weights of all layers, except the last layer. Replace last layer of TF SavedModel with new layer based on categories of new dataset. Re-train the weights for last layer only. | No |
| Freeze weights upto any arbitrary Nth layer (given layer name). Replace final layer of TF SavedModel with new layer based on categories of new dataset. Re-train the weights from (N+1) th layer to the last layer. | No |

@zeahmed @JRAlexander