As said in #1895, we need faster parsing of UTF-16 escaping sequences, this PR is the answer.
It makes the runtime of the `large_escaped` benchmark go down from ~5s to ~3.5s, and with valgrind, from 26GIr to 20GIr
Note: based on #1886, only the 4 last commits are of interest here
Pull-Request: #1896
Reviewed-by: Jean Privat <jean@pryen.org>