Preserve JDK decoding with fewer copies

Binary-to-STRING casts and UTF-8 DECODE use the same native decoder. Released Flink constructs a Java String with the UTF-8 charset, whose replacement grouping differs from Rust's lossy UTF-8 conversion. For example, ED A0 80 becomes one U+FFFD. The native path preserves that grouping; this is separate from raw FROM_BASE64 STRING transport.

Well-formed Arrow Binary input reuses its value, offset and validity buffers as Utf8. FixedSizeBinary constructs offsets and shares its sliced bytes and NULLs. For malformed input, the decoder copies the validated prefix in one operation, then handles malformed units according to the JDK contract. Long ASCII runs use block checks. Runs of ASCII and independently invalid bytes are encoded in a bounded stack buffer, reducing three-byte builder writes. Decoding writes into the final Arrow value buffer, avoiding a temporary String and a second copy. Other charsets retain their existing conversion paths.

Modern bounded binary-to-character casts also reuse decoded buffers when every row already satisfies the target width. Otherwise they write final values and offsets directly, keeping the input validity. VARCHAR truncation checks only the required ASCII prefix before falling back to UTF-16 unit counting; CHAR additionally pads with spaces. A cut through a supplementary pair writes the JDK's ? replacement for the remaining high surrogate. Consumers that need the original Java String keep the generated expression, as documented on the Calc coverage page.

This follows Comet's valid-UTF-8 borrowing, explicit JDK replacement rules and direct Arrow-builder writing structure. Flink's released results define the contract. No JNI interface, dependency or inter-operator string representation changes. The Calc coverage page documents retained STRING identity restrictions in issue #493.

Correctness checks compare complete output against released generated casts, including malformed continuation prefixes, surrogate encodings, NULLs, fixed padding, slices and empty batches. Repeated batches must retain independent output ownership. Decoder tests exercise long valid runs on both sides of a malformed unit and stack-buffer boundaries.

Measurements and the reproducible stock/previous/native comparison are recorded in the scalar benchmark report.