Add support for custom missing value in ValueMappingTransformer.

The `ValueMappingTransformer` currently handles missing values. It does so by using the default value for the value type.

https://github.com/dotnet/machinelearning/blob/b4c10662706589426a0a870304e3088dec8077ac/src/Microsoft.ML.Data/Transforms/ValueMapping.cs#L807

In some cases, it is desirable to map the value that is missing in the dictionary with the user specified value. For example, when using the following text classification model from TensorFlow
https://github.com/tensorflow/models/tree/master/research/sentiment_analysis

, an `out-of-vocabulary` word (missing value) is mapped to integer `3`. This cannot be done currently using `ValueMappingTransformer` in ML.NET.