Add a new HTML rendering pipeline

[ehuss]

This rewrites the HTML rendering pipeline to use a tree data structure, and implements a custom HTML serializer. The intent is to make it easier to make changes and to manipulate the output. This should make some future changes much easier.

This is a large change, but I'll try to briefly summarize what's changing:

• All of the HTML rendering support has been moved out of mdbook-markdown into mdbook-html. For now, all of the API surface is private, though we may consider ways to safely expose it in the future.
• Instead of using pulldown-cmark's html serializer, this takes the pulldown-cmark events and translates them into a tree data structure (using the ego-tree crate to define the tree). See tree.rs.
• HTML in the markdown document is parsed using html5ever, and then lives inside the same tree data structure. See tokenizer.rs.
• Transformations are then applied to the tree data structure. For example, adding header links or hiding code lines.
• Serialization is a simple process of writing out the nodes to a string. See serialize.rs.
• The search indexer works on the tree structure instead of re-rendering every chapter twice. See html_handlebars/search.rs.
• The print page now takes a very different approach of taking the same tree structure built for rendering the chapters, and applies transformations to it. This avoid re-parsing everything again. See print.rs.
  • I changed the linking behavior so that links on the print page link to items on the print page instead of outside the print page.
• There are a variety of small changes to how it serializes as can be seen in the changes to the tests. Some highlights:
  • Code blocks no longer have a second layer of <pre> tags wrapping it.
  • Fixed a minor issue where a rust code block with a specific edition was having the wrong classes when there was a default edition.
• Drops the ammonia dependency, which significantly reduces the number of dependencies. It was only being used for a very minor task, and we can handle it much more easily now.
• Drops pretty_assertions, they are no longer used (mostly being migrated to the testsuite).

There's obviously a lot of risk trying to parse everything to such a low level, but I think the benefits are worth it. Also, the API isn't super ergonomic compared to say javascript (there are no selectors), but it works well enough so far.

I have not run this through rigorous benchmarking, but it does have a very noticeable performance improvement, especially in a debug build.

I expect in the future that we'll want to expose some kind of integration with extensions so they have access to this tree structure (or some kind of tree structure).

Closes #1736

[max-heller]

This is exciting to see! In mdbook-pandoc, I do something very similar, I'd certainly be interested in seeing how much of this functionality could be shared between mdbook and backends.