Allow me to opt out of replacing missing values (missing_value_replacement=None)

[npatki]

Problem Description

A number of the RDT transformers have a missing_value_replacement parameter that describes how to replace missing values in the forward transform. There are a variety of strategies that include replacement with a random value ('random'), replacing with the mean of the data ('mean'), etc. This was done under the assumption that the downstream software does not accept missing values.

However, I may have some use cases where my downstream software does handle missing values. In this case, I do not want the RDTs to do anything to the missing values. In such cases, we should offer missing_value_replacement=None as an option.

Note that we used to have this option. Just that it is listed as deprecated for all the transformers. We should un-deprecated it because there are valid use cases that need this.

Expected behavior

Reinstate the None option for the missing_value_replacement parameter, wherever it exists. If this is passed in, then during the forward transform: do not replace the missing values with anything. Just pass the missing values along.

• Note that the default value for missing_value_replacement should not change; in most cases, it is 'random'. The user would need to explicitly pass in missing_value_replacement=None to access this new functionality.

For the reverse transform things can be a bit trickier. Ideally the missing_value_generation parameter is supposed to tell us how to recreate missing values. But it the passed-in data already contains missing values, then it can mess up generation. Here's what we can do:

If the missing_value_generation is not None and the passed in data contains some null values, then don't do anything to the missing values. Instead, show the user a warning.

Warning: The 'missing_value_generation' parameter is set to 'from_column' but the data already contains 
missing values. Missing value generation will be skipped.

Additional context

This change should probably be done to a base class. It would ultimately affect the following transformers:

• numerical: ClusterBasedNormalizer, FloatFormatter, GaussianNormalizer, XGaussianNormalizer
• categorical: BinaryEncoder
• datetime: OptimizedTimestampEncoder, UnixTimestampEncoder