Improve the default HTML parser and document normalizer extension

[gschlager]

While reviewing an HTML migration adapter a colleague built on Markbridge, I found several gaps in the default HTML parser, the renderer, and the docs. This PR closes them. This is the first of two PRs from that review — the deeper design changes (ancestry matching for AST subclasses and the AST::Code block flag) are stacked on top in #72.

• The HTML handler registry had no handler for <h1>–<h6>, so headings from HTML input lost their structure and only their text survived. There is now a HeadingHandler that stores the level in the existing AST::Heading node; the renderer side already worked through the auto-registered HeadingTag.
• RawHandler used the whole class attribute as the code language, so styling classes like hljs ended up on the fence line, and it never looked at the <code> child inside <pre>, where the CommonMark convention puts the language-* class. The language now comes from the first of: a language-* class on the element or its direct <code> child, the lang attribute, or a lone class used as-is. The result must be a single clean token, so it is always safe on the fence line.
• Empty elements left junk in the output: <code></code> rendered as a bare `` pair, an empty details block as an empty [details] shell, and an empty spoiler as `[spoiler][/spoiler]`. All three now render to nothing.
• The html_mode contract check now ships with the gem: require "markbridge/rspec" provides the shared example "an html_mode safe tag", so adapter authors can put their custom tags under the same check we run against the built-in tags. The check itself moved into Renderers::Discourse::HtmlBlockSafety and our contract spec runs through the shared example.
• The HTML parser docs no longer show a handler signature with a processor: parameter that does not exist, a RawHandler.new call without its required argument, or lambda handlers, which the parser does not support. The docs site build output is now in .gitignore.

Behavior changes to be aware of: an explicit lang attribute now wins over a lone class without language- prefix, because a class can be pure styling while lang is always set on purpose. And empty code/details/spoiler elements disappear from the output instead of leaving empty shells.