perf: avoid per-row copy in Spark hex byte encoding

[andygrove]

Which issue does this PR close?

N/A

Rationale for this change

The Spark hex function's byte-encoding path (hex_encode_bytes) filled a reusable per-row Vec and then copied it into a StringBuilder for every row. Since each input byte maps to exactly two hex characters, the output can be written once directly into a single value buffer, eliminating the per-row copy.

What changes are included in this PR?

Rewrote hex_encode_bytes to write hex digits directly into one growing value buffer with manually-tracked offsets and a NullBufferBuilder, constructing the StringArray via new_unchecked (the buffer contains only ASCII hex digits, so the existing no-validation behavior is preserved). The 256-entry lookup table and all other hex paths are unchanged.

Are these changes tested?

Existing tests (function::math::hex) pass unchanged.

Benchmark (datafusion/spark/benches/hex.rs):

• hex_binary/1024·4096·8192: ~13.8% / 14.3% / 13.9% faster
• hex_utf8/1024·4096·8192: ~14.1% / 14.9% / 15.0% faster

Are there any user-facing changes?

No