Allow smart "merging" of schema for multi-file schema inference

[seandavi]

Use case

When performing schema inference on multiple files, there is an opportunity to be more flexible by having clickhouse "union_by_name" the schemata to allow these schema variants to load as a unified table. This would be similar to read_parquet('*.parquet', union_by_name=True) to allow schema inference to simply add new columns when they are not present in all files?

The same logic could apply to CSV, TSV, jsonlines, etc., but my use case is for Parquet, so feel free to scope as you see fit.