Defer visible/invisible HTML rendering until actually requested

[bigio]

rendered() unconditionally computed get_rendered_text() three times per HTML part (full, visible-only, invisible-only), each doing a full mask-and-whitespace-cleanup pass over the text. Keep the HTML parser object around instead and only compute the visible/invisible variants inside visible_rendered()/invisible_rendered() when something actually calls them, since not every ruleset uses invisible-text rules.

[fkoyer]

I like the idea of doing things lazily to improve performance. I also like the idea of moving towards making Node.pm type-agnostic. In other words, I was thinking about moving all HTML handling into the HTML handler at some point in the near future. Currently that would mean moving lines 814–892 of Node::rendered() into Handler::HTML. This change makes that slightly more complicated because both visible_rendered and invisible_rendered need to know about html_obj. Your change works because, right now, text/html is the only type that emits invisible text. I like to keep things pure in case someday there's another part type that emits invisible text (I don't know what that would be). But I don't dispute that this change has a slight performance advantage, if there are no invisible text rules.

One more thing to note, bayes_token_sources defaults to "header visible invisible uri" so if Bayes is enabled, both visible_rendered() and invisible_rendered() fire on every HTML part and both deferred passes run anyway.

[bigio]

We could revisit this once you moved the code into the HTML handler, I will keep the PR open as a reminder.

[fkoyer]

One possible idea: check invisible_rendered first, if it returns nothing then set visible_rendered = full rendered. That would save one pass if there's no invisible text.