Summary

This PR removes the need for an Rc to implement interned by using the Buffer as an interner

• Create an ID for the interned content
• Wrap the content in StartInterned and EndInterned tag with the corresponding ID and write the content directly to the target Buffer.
• Write a Reference(id) node to reference to the interned content.

The Printer skips interned content but writes it to an internal Index. The Printer uses the internal Index to resolve the Interned content when it sees a Reference. The fact that ids tend to be monotonic allows the use of a Vec as an Index similar to Group ids

The downside of this approach is that it is more complicated to reason about and it is no longer possible to inspect the content of Reference without tracking all interned content. Another downside is that the IR is slightly misleading because you see Interned content and would assume it gets printed in that place, which is not true. For example, the interned content appears before the best_fitting, but the content only gets printed inside the best-fitting variants. Moving the Interned content into the first best fitting variant doesn't work because that content only gets printed conditionally, meaning the Printer would then not always see the Interned content, failing to resolve the Reference.

Alternatives

One alternative that I explored is to use bumpallo instead and use it to allocate BestFitting vectors, dynamic strings, and Rcs. This could improve performance even more because it eliminates the 3-4% drop-time too. However, it comes at the downside that Buffer requires a new 'elements lifetime that needs to be added everywhere (Format, Formatter, Buffer, FormatNode, FormatNodeRule....). I gave up at some point because it made the API significantly more complicated.

Test Plan

cargo test