This PR simplifies json_string_unescape by having convert_UTF32_to_UTF8 write directly to the output buffer. This saves a MEMCPY as the unescape_len isn't known at a compile time so it isn't optimized away. At least that's what it seems based on the samply profile:

I'm pretty sure this doesn't introduce any potential out-of-bounds write as the current code in master unconditionally writes unescape_len bytes to buffer whereas this branch passes buffer directly to convert_UTF32_to_UTF8 which writes the same unescape_len bytes to the buffer.

Before

After

Benchmarking on my M1 Macbook Air shows a 2% improvement when parsing activitypub.json depending on run.

== Parsing activitypub.json (58160 bytes)
ruby 3.4.8 (2025-12-17 revision 995b59f666) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
               after     1.059k i/100ms
Calculating -------------------------------------
               after     10.891k (± 1.0%) i/s   (91.82 μs/i) -     55.068k in   5.056759s

Comparison:
              before:    10694.4 i/s
               after:    10891.0 i/s - 1.02x  faster