Experimental native Parquet readers unpack dictionaries by default

[mbutrovich]

What is the problem the feature request solves?

When using native_datafusion or native_iceberg_compat Parquet readers based on DataFusion's DataSourceExec, the schemas that Comet passes in result in dictionaries being unpacked immediately.

Describe the potential solution

Arrow-rs will use a provided schema as a hint, and in the case of dictionary encoded columns, preserve the encoding:
https://github.com/apache/arrow-rs/blob/880be2f0a0b9675d8b42206e70543472a58792aa/parquet/src/arrow/schema/primitive.rs#L91

The challenge is similar to int96, where the native side doesn't really have the Parquet schema when generating the DataSourceExec. We'd either need to pass this from early on the Spark side when the schema is first read, or add a coercion rule to DataFusion.

Additional context

No response