Lone high surrogate "\ud800" is decoded as U+0000 instead of an error

[ddxd]

A high surrogate escape that is not followed by a low-surrogate escape is decoded as U+0000 instead of producing an error, so invalid input silently turns into different data.

Reproduce (simd-json 0.18.1)

let mut v = br#""\ud800""#.to_vec();
let t = simd_json::to_tape(&mut v).unwrap();   // Ok
assert_eq!(t.0[0], simd_json::Node::String("\0")); // "\ud800" became "\0"

The same happens for "\ud800x", ["\udbff"], "\ud800\n", and for a pair whose second escape has invalid hex digits ("\ud800\uzzzz"). A lone low surrogate ("\udc00") and "\ud800A" are correctly rejected. It reproduces through to_tape, to_borrowed_value and serde, on every SIMD implementation, because they all go through get_unicode_codepoint.

Cause

In src/stringparse.rs, get_unicode_codepoint returns Ok((0, src_offset)) in both of those cases. That 0 is a leftover sentinel. When this code returned the number of bytes written (handle_unicode_codepoint before 29d9998), 0 meant "failed", and callers checked if o == 0 { return Err(...) } (those checks are still in the avx2/sse42/neon/simd128/portable parse_str). Since 29d9998 ("Implement std::simd portable") the function returns the code point instead, so the 0 means U+0000, codepoint_to_utf8 writes a NUL byte, and the o == 0 checks can no longer fire.

Expected

An error (InvalidUnicodeCodepoint), as for a lone low surrogate, and as serde_json, sonic-rs and orjson do.

I have a fix with a regression test and will open a PR referencing this issue.