It seems we have a very long-standing typo in the journal replay magic number
unescaping code that basically means it will cause data corruption if we ever
actually use it. I think this is a candidate for the stable tree.

The following script, run on an FS mounted with data=journal, may be helpful in
reproducing and testing this bug:

#!/bin/bash

rm -rf test-dir && mkdir test-dir

for (( i=0; i != 100; i++ )); do
	printf "\xc0\x3b\x39\x98" > test-dir/$RANDOM
done

while /bin/true; do
	printf "\xc0\x3b\x39\x98" > test-dir/$RANDOM
	rm test-dir/`ls test-dir | sort -R | head -n1`
done