DType is not sufficient to correctly infer and export arrow schemas

[AdamGS]

For ParquetVariant (and potentially other types), the array's metadata is required to accurately provide an Arrow Field. The physical layout of the array itself influences the Arrow-side data type.

Parquet solves this in two ways:

1. Storing the schema as file metadata.
2. Parquet is structured as a tree of types, so Variant columns have children that can be directly inspected to infer the schema correctly.