[Rich-T-kid]

This is stacked on #26973; the final two commits are the change in this PR.

When the table schema uses Dictionary(Int32, Utf8) but the physical Parquet schema is LargeBinary, the reader currently emits plain values and DecoderProjection dictionary-encodes every row. Preserve the physical value type in the reader schema instead, so arrow-rs decodes dictionary keys directly and the projection casts only the distinct dictionary values. The same rule handles Utf8/LargeUtf8 offset-width mismatches.

The Q23 profile no longer contains pack_byte_to_dictionary or GenericByteDictionaryBuilder::append_array; it now enters parquet-rs DictionaryDecoder directly. A controlled ClickBench run is pending on #26964.

Validated with cargo test -p datafusion-datasource-parquet --lib, package clippy with all targets/features, and the repository pre-push checks.

Closes #26976.

[Rich-T-kid]

Hosted ClickBench result: #26964 (comment)

With dictionary mode enabled on both sides, this change reduced total time from 22.22s to 18.87s (-15.1%) with 3 faster, 0 slower, and 40 unchanged queries in the primary comparison. Q10 improved 198.86ms -> 65.91ms (3.02x), Q11 211.87ms -> 76.34ms (2.78x), and Q23 6146.32ms -> 2992.38ms (2.05x). Peak RSS fell from 12.6 GiB to 11.3 GiB.