Improve implementation of ``NullDataCheck``

Follow up on https://github.com/alteryx/evalml/pull/3182 based on @freddyaboulton's comment:

I think we can improve this implementation. Right now we do two scans of the data to determine the highly null columns and the columns with any nulls. Doing so also involves creating a new copy of the data without NaturalLanguage/DateTime.

I think if we refactor get_null_column_information to return {col_name: % nulls} we only have to do one scan without creating another copy to determine which are highly null and which are null.

Originally posted by @freddyaboulton in https://github.com/alteryx/evalml/pull/3182#discussion_r789977564