Tracking issue for better Enum / Categorical support in Polars Parquet

[coastalwhite]

At the moment, several parts of Enum's and Categorical's are broken in the conjunction with Parquet reading and writing. This mostly stems from the fact that Polars does not use the arrow Dictionary type to store data contained in enums and categoricals. Instead, it chooses to store them in primitive u32 arrays with the categories being stored in the DataType. This causes us to need to make a small translation step when writing handling parquet for interoperability with other parquet readers.

Some things that are currently broken or less than ideal.

• You need to load at least one row to get back your categories. (write_parquet does not preserve enum categories for empty data frames #20083)
• Enum's and Categoricals have to be disabled from most parametric tests, which has caused bugs in production before.
• Enum's and Categoricals are almost randomly swapped when reading and writing (Enum type becomes a Categorical type when file is read with scan_parquet #19389)
• Enum's and Categoricals are not supported in predicate filtering (Parquet scanner doesn't do predicate pushdown for categoricals/enums #18868)
• Min and max values are too pessimistic (Parquet writer statistics for categorical/enum values have overall min/max instead of row_group min/max #18867)
• Dictionary types written by other writers with PLAIN encoding or different categories between row groups are not supported by Polars and may produce invalid results (Parquet writer statistics for categorical/enum values have overall min/max instead of row_group min/max #18867)

When these issues are fixed, working with Enum's and Categorical's in polars parquet should be a lot more doable.