Union nested Parquet leaves across projection accesses

When one projected nested column was consumed through multiple narrowing casts, or through both a narrowing cast and get_field, the Parquet read plan fell back to reading the entire root column even though the required leaf set was computable.

Accumulated and deduplicated leaf offsets across narrowing casts and get_field accesses, derived the emitted Arrow type from the merged leaf set, and retained full-read fallback behavior for unsupported partial wrapper shapes.

• Reduces unnecessary Parquet I/O for queries with multiple nested accesses to the same root column.
• Preserves existing SQL/API behavior while improving scan efficiency for supported nested projection patterns.
• Keeps correctness by unioning all required leaves rather than clipping to one consumer or reading the whole root.

• Added coverage for disjoint, overlapping, and repeated casts, cast plus get_field, and casts diverging below List<Struct>.
• SQL logic tests verify returned values and confirm merged projections scan fewer bytes than a full-root read.
• Ran datasource-parquet library tests, parquet_nested_schema_pruning sqllogictests, formatting, and strict clippy for the changed package.