This avoids having to overestimate the md5 context size, which
isn't known beforehand, allowing us to use the new allocate functions
instead.
Signed-off-by: Martin Storsjö <martin@martin.st>