mark_safe and mark_for_escaping should have been kept similar.
On Python 2 this change has no effect. On Python 3 it fixes the use case
shown in the regression test for mark_for_escaping, which used to raise
a TypeError. The regression test for mark_safe is just for completeness.
Backport of 5c5eb5fe from master.
This reverts commit 2ee447fb5f.
That commit introduced a regression (#21882) and didn't really
do what it was supposed to: while it did delay the evaluation
of lazy objects passed to mark_safe(), they weren't actually
marked as such so they could end up being escaped twice.
Refs #21882.
The idea is that if an object implements __html__ which returns a string this is
used as HTML representation (eg: on escaping). If the object is a str or unicode
subclass and returns itself the object is a safe string type.
This is an updated patch based on jbalogh and ivank patches.
