Refactor benchmarks

[nicoburns]

Objective

• Deduplicate benchmarking code
• Make it easier to add new benchmarks with different style setups
• Enable a future extension which benchmarks older versions of Taffy directly against the latest version (head-to-head like we do for yoga).

Changes made

• Abstract tree creation behind a trait
• Implement trait for Yoga and Taffy

Notes

This change seems to have changed some of the benchmark results quite a bit, which I believe is due to a change in node creation order, which interacts with the randomly generated node styles. This suggests that we ought to make sure that the Yoga and Taffy trees are being created in exactly the same order (so that exactly the same tree is created). I'd also like to play around with different random seeds and see how this affect things.

Feedback wanted

• Does this seem like a good direction? It does complicate the benchmarking code a bit by introducing abstraction where previously there was simple concrete code.

[alice-i-cecile]

I like this direction, and I think the abstraction pays off. Definitely agree that we need to ensure we're using exactly the same trees.

[nicoburns]

Marking this as ready for review / merging.

I'm seeing much more consistent results with this PR now. I have checked, and it seems to me that the tree structures are now the same (I think I must have fixed this ages ago). I've also tried out different random seeds, and it doesn't seem to make that much difference (+-5%, and movement in the same direction for both Taffy and Yoga).