Implements the filtering logic for
`suggestion:-x+y` syntax.

The spec needed to be updated because in some cases the resulting
Markdown is slightly different, though equally valid.

Excludes a few filters that require more work:

* lib/banzai/filter/image_lazy_load_filter_spec.rb
* lib/banzai/filter/syntax_highlight_filter_spec.rb
* lib/banzai/filter/table_of_contents_filter_spec.rb

Part of #47424

This reverts commit 96bef54154e669f9a3e92c3a4bc76c0be3a52e48.

Also break up a long line, just 'cause.

Fixes: https://gitlab.com/gitlab-org/gitlab-ce/issues/18096

IssueReferenceFilter will end up processing internal issue references
when a project uses an external issues tracker while still using
internal issue references (in the form of `#\d+`). This commit ensures
that these links are rendered as external issue links, regardless of
whether the project one currently views uses an internal or external
issues tracker.

Fixes gitlab-org/gitlab-ce#19036, gitlab-com/performance#16

This reverts commit 3e991230.

# Conflicts:
#	app/models/project.rb

This method returns an Array of the HTML nodes as yielded by
ReferenceFilter#each_node. The method's return value is memoized to
allow multiple calls without having to re-query the input document.

This changes ReferenceFilter#each_node so that when it's called without
a block an Enumerator is returned.

This splits the Markdown rendering and reference extraction phases into
two distinct code bases. The reference extraction phase no longer relies
on the html-pipeline Gem (and any related code) and allows for
extracting of references from multiple HTML nodes in a single pass. This
means that if you want to extract user references from 200 comments you
no longer need to run 200 times N number of queries, instead only a
handful of queries may be needed.

These filters now use a single iteration over all the document nodes
instead of multiple ones. This in turn allows variables to be re-used
(e.g. links only have to be unescaped once). Combined with some other
refactoring this can drastically reduce render timings.

We're trying to avoid circular dependency errors.

Closes #6077

This is a tad faster than letting Nokogiri figure out whether it should
evaluate the query as CSS or XPath and then actually evaluating it.