perf(leptos): erase nested views and event callbacks - #4879

Summary

• Erase row views and callbacks inside For and ForEnumerate.
• Erase children inside Suspense, Transition, and ErrorBoundary, and box the ErrorBoundary fallback callback.
• Erase child views inside Show and ShowLet, and box event callbacks inside on when erase_components is enabled. on_target reuses its existing callback box.
• Keep these conversions behind erase_components so the normal configuration retains its existing view and callback types.

Motivation

erase_components erases component return types, but nested views and event callbacks remain concrete. My codebase has about 1 million lines of Rust; a source scan found about 3,200 #[component] definitions and 11,200 view! calls, including roughly 450 keyed lists, 3,600 Show/ShowLet uses, 1,500 Suspense/Transition uses, and 3,200 event attributes. Erasing these internal types aims to shorten my development compile loop. Callback boxing adds allocation and dynamic dispatch; the probe below measures compile time, not runtime cost.

Compile time probe

I compiled generated Rust binaries against leptos_0.9 with --cfg erase_components on both sides. Every binary references all its view! modules. The keyed and boundary workloads each contain 24 modules with different view and closure types; the varied workload has 24 distinct structures; the combined workload has 72 modules drawn from keyed, boundary, and effect views. The mixed workload has 24 modules with different HTML layouts, each containing Show, ShowLet, and six event handlers. Its baseline already includes the For and boundary changes, whereas the first four rows compare those changes against their original implementations.

Each comparison used five alternating trials on rustc 1.100.0-nightly, warm dependencies, development code generation, CARGO_INCREMENTAL=0, an empty RUSTC_WRAPPER, and fresh rustc metadata. The figures are median rustc total seconds from -Z time-passes; WebAssembly used the hydrate feature.

Macro expansion was nearly unchanged in the varied workload (0.044 s original, 0.045 s optimized), so that gain came after expansion. The optimized mixed workload was faster in all five paired runs for each target. These local fixtures are not part of the PR, and the figures exclude dependency rebuilds, bundling, and a whole application build.

Verification

• cargo fmt --check --package leptos --package tachys passed.
• A local generated fixture passed SSR rendering and headless Firefox click tests with the optimization enabled and disabled, including typed-target handlers.