fix: avoid hex decoder UTF-8 slicing panic

[vycdev2]

Summary

• decode hex input from raw bytes instead of slicing UTF-8 strings at byte offsets
• preserve existing odd-length, invalid-hex, and invalid-UTF-8 errors
• add direct comptime and registered runtime-failure regressions for non-ASCII input

Verification

• cargo fmt --all -- --check
• cargo check
• cargo build
• cargo test -q -p jett_comptime builtin_encoding_hex_decode_rejects_non_ascii_without_panicking
• cargo test -q -p jett_driver encoding_hex_decode_rejects_non_ascii_without_panicking
• cargo run -q -p jett_cli -- format --check tests/runtime_fail/encoding_hex_decode_non_ascii.jett
• cargo test -q (433 driver tests passed)
• git diff --check

Risk

• Low: focused decoder implementation change; no API or error-message changes.
• The related encoding contract issue is design: define encoding representations and failure contract #71; this PR supplies the runtime panic fix.

This was generated by an AI agent (vycdev2). Please verify any changes before merging or applying.

[vycdev]

The byte-pair hex decoder correctly removes the UTF-8 slicing panic and its tests passed, but this branch now conflicts with current main in the shared interpreter/fixture area. Please rebase, preserve the already-merged encoding and checked-math regressions, and rerun comptime/driver tests.