Add a config to set the footnote separator. Also remove rel/rev as they aren't really compatible with HTML5 and we already have classes set for refs and backrefs.

Fixes 723.

Fixes #412 and #493.  First we parse footnote content as
its own document avoid quirks with using li as a parent. Second, we
surround placeholders with STX and ETX to prevent them from interfering
with inline parsing; this is also consistent with how placeholders are
used everywhere else in Python Markdown.

Track when we find duplicate footnote references and create unique ids
for them.  Then add an additional tree-processor after inline to go
back and update the footnotes with additional back references that link
to the duplicate footnote references. Fixes #468.

As HTML5 has depreciated use of `rev=anything` and `rel=footnotes`, they are no
longer inlcuded in the output when the output_format is set to HTML5. Note that
if someone successful registers a spec for `rel=footnotes` in the future (as
a microformat), then that could be considered valid. But until that happens,
it is invlaid to use in HTML5. Therefore, we remove it from the output (when
outputing HTML% only).

As an alternative, two new classes are set (in all output_formats). On the link
to the footnote (where `rel=footnotes` was used), we set `class=footnote-ref`
and on the backlink (where `rev=footnote` was used), we set
`class=footnote-backref`.

Also updated the tests to reflect to the new classes in the output.