we rewrite the dump code completely, stealing bits from
safe_shell_exec() for the DB credentials, to directly do the search
and replace of certain patterns ourselves. this simplifies the
pipeline and should also improve performance (because we avoid forking
sed twice), unless some memory leaks exist in PHP (which is quite
possible). we process the dump one line at a time so memory usage
shouldn't be higher.

the previous code would not catch errors from mysqldump because the
"sed" pipeline would succeed and therefore hide error from earlier in
the pipeline. we would need "set -o pipefail" in bash to workaround
this issue, but that is not portable and, let's be honest, there
already enough crap here as it is to warrant a cleanup.

previous attempt at fixing this issue (#2098289) are sometimes
unsuccesful because some dumps fail later than 1024 bytes. we still
keep that check just in case.

the bug (#1324466) was introduced in 571b63de, present since the 1.1
release.

this should be refactored to make safe_shell_exec() allow for a hook
to modify the dumpfile inline, but for now i just wanted to port a
working prototype in place.