Huge regression in IsolateJson benchmark

[askeksa-google]

The change to write UTF-8 decoded characters directly into the result strings resulted in a 8x-20x slowdown of the IsolateJson.Decode* benchmarks on ARMv8.

It looks like the central decode16 method is compiled through OSR for every call to it (200 calls in total when running all of those benchmarks). Deoptimization tracing indicates that the method is never deoptimized.