Optimize C++ ProtoJSON by scanning ahead on already-available input chunks where we would otherwise scan by always peeking one char at a time.

This improves our synthetic parse benchmark by about 20%; especially since JSON parsing is extremely heavy on looking for quoted strings (all keys and a significant number of the values) we otherwise burn a lot of time on peeking one character at a time.

PiperOrigin-RevId: 911516316