Refactor frame metadata into tags

[laurmaedje]

This replaces the MetaElem with a new TagElem that fills a similar role, but works a bit differently. The purpose of both is to hold elements that will become available through introspection. However, the MetaElem did this via the style chain, by attaching a style to the full subtree of content resulting from a locatable element, which resulted in lots of redundant FrameItem::Meta(Meta::Elem(_), ..) in the frames.

The new TagElem is generated exactly once per locatable element and holds the element directly. The challenge here is that the tag still needs to stick to the element so that if e.g. a heading moves to another page, the page number in the outline is still correct. To that extent, the flow layouter has gained a new pending_tags array and the tags are only flushed into the frames when something proper is layouted. Overall, the new mechanism is much more efficient. The old thing was actually so bad that the famous thesis benchmark improves by ~14% with this PR.

The old FrameItem::Meta was also used for hiding and links. These usages have been replaced by a completely internal mechanism for hiding and a new top-level FrameItem::Link for links.

This PR is only a first small step in an incremental refactor of realization and flow layout. The TagElem and pending_tags aren't necessarily the best way to implement tags in realization and layout, rather they are the least intrusive ones in the current setup. Things will change further as the refactoring progresses. The exact semantics of how tags will stick to elements are also still somewhat in-flux and will likely a change a bit in later refactor phases. These rules affect the results of .page() and .position() calls on locations. With this PR, things should mostly be the same as on main, though there will be edge cases where things changed a bit. They might change again with further refactoring.