many_buttons performance regression

[DGriffin91]

Bevy version

This regression was introduced in Remove custom rounding (#16097) at 05d9068

The commit just before it: af27907 is faster.

cargo run --release --example many_buttons
Before the #16097 PR: 17.6ms, after: 21.7ms.

Windows 10, 7950x, RTX3060, Vulkan

After the PR (05d9068)

Before the PR (af27907)

vtune with cargo run --profile release-with-debug --example many_buttons

[profile.release-with-debug]
inherits = "release"
debug = true

[ickshonpe]

Using Taffy's custom rounding is quite a bit slower as it walks the node tree an extra time after generating the UI layout. The less accurate custom rounding code that was removed applied the rounding during the walk to update all the nodes.

Maybe we could copy Taffy's rounding into bevy_ui_layout system, but it's a bit fragile and overcomplicated already which is why I didn't do that in #16097.