map Iceberg Binary to Arrow Binary instead of LargeBinary

[parthchandra]

Follow up from #2668

Summary

iceberg-rust currently maps PrimitiveType::Binary to DataType::LargeBinary in the Arrow type conversion (type_to_arrow_type). This is inconsistent with:

1. Java's implementation — ArrowSchemaUtil.java#L134 maps Iceberg Binary to Arrow Binary.
2. parquet-rs behavior — parquet-rs reads binary columns from Parquet files as DataType::Binary, not LargeBinary.

This mismatch causes unnecessary casts when reading Binary columns and creates inconsistency with the Java ecosystem.

Scope

• Change PrimitiveType::Binary mapping from DataType::LargeBinary to DataType::Binary in crates/iceberg/src/arrow/schema.rs
• Update create_primitive_array_repeated in crates/iceberg/src/arrow/value.rs — the existing DataType::Binary arm already handles this, but the LargeBinary arm added in feat: add support for _partition metadata column #2668 can be removed or kept as a fallback
• Audit all places that construct or expect LargeBinary arrays for Iceberg Binary columns
• Update tests that assert LargeBinary output

Context

Raised in review of #2668 by @advancedxy. The LargeBinary arm in create_primitive_array_repeated was added to handle the current mapping for _partition columns partitioned by Binary. Once this mapping is corrected, that arm becomes unnecessary (but harmless to keep as a safety net).

This is a broader change that affects all Binary column reads, not just _partition.