fix(smx): gif pack option display and CanBeEmpty fallback semantics

[fchorney]

Rebased onto current main. #617 is stacked on this branch, so this one lands first.

Two fixes for the gif-pack bug reports against v0.4.2004.

1. Selected gif pack shows as Default when re-entering the options screen

Selecting a pad or judgement gif pack on the StepManiaX options page appears not to stick: re-entering the options screen always shows Default checked, regardless of what was selected.

The saved value is actually fine the whole time: the config, the ini round-trip, and the pad lighting all keep the chosen pack. Only the displayed row selection resets.

Cause: the SmxBgPack / SmxJudgePack rows are defined with a single static placeholder choice; their real choice list (Default plus the discovered packs) is provided dynamically by the options layout. The screen-init sync used set_choice_by_id, which clamps the incoming index against the static row definition's choice list:

// src/screens/options/row.rs
let max_idx = rows[pos].choices.len().saturating_sub(1); // placeholder len = 1 -> max_idx = 0
*slot = idx.min(max_idx);                                // always clamped to 0 = "Default"

Fix: write the index directly via get_choice_by_id_mut, the same pattern already used by the other dynamically-populated rows (DefaultNoteSkin, SoftwareRendererThreads). The remaining dynamic rows (display mode/resolution/refresh rate, sound device) already use direct writes; the per-player pack rows in Player Options bake their full choice list into the row and are not affected.

2. CanBeEmpty roles still slid over to the pack's default gif

A pack that supplies all of its own gifs except gameplay and declares CanBeEmpty = "gameplay" expects no background during gameplay. The registry honored the declaration for the pack-level chain (no Fallback pack, no common), but the app's screen resolution then fell back to the default role, so the pack's own default gif showed during gameplay anyway.

Fix: the role-candidate walk (grade-specific results roles, then the screen role, then default) now stops dead at any candidate the selected pack declares under CanBeEmpty without supplying: that event shows no animation at all. CanBeEmpty now means what it says: this name never has an animation, no matter which fallback would have provided one.

Details:

• A gif the pack supplies still wins over its own declaration (unchanged).
• Per-song / per-pack scoped gifs sit above pack policy and are unaffected (they are an override, not a fallback).
• New registry API GifRegistry::background_declared_empty exposes the declared-empty state (the existing background() lookup cannot distinguish "empty by declaration" from "missing, keep falling back"), with a unit test.
• docs/stepmaniax.md updated: the resolution chain, the grade-chain section, the CanBeEmpty table row, and the "when to use" example now describe the stop-dead semantics. Also corrected the section that claimed an opted-out role shows solid black with the game holding LED ownership at all times: actual behavior is solid black during gameplay (the game still owns the LEDs for judgement effects), and outside gameplay an empty pad is handed back to the pad firmware's built-in lighting.

Testing

• cargo test -p deadsync-smx: 76 passed (includes the new background_declared_empty test).
• cargo build clean.
• Manual (options display fix): select a pack on the SMX options page, leave the screen, re-enter; the row now shows the selected pack instead of Default.

[fchorney]

Heads up: main currently doesn't build on macOS

Not caused by this PR, and not fixed by it. Fix is up as #618, which branches straight off main and is independent of this one.

I hit these while rebasing onto main (146d8241). Both are inside #[cfg(target_os = "macos")] blocks, which is presumably why CI is green.

1. crates/deadlib-renderer/src/window_size.rs:12 (from de4a1c5b)

const fn macos_opengl_low_dpi(backend_type: BackendType, high_dpi: bool) -> bool {
    backend_type == BackendType::OpenGL && !high_dpi
}

error[E0015]: cannot call non-const operator in constant functions

== on BackendType goes through PartialEq::eq, which isn't callable in a const fn on stable (it needs const_trait_impl). #618 uses matches!(backend_type, BackendType::OpenGL) instead: pattern matching calls no trait method, so the function stays const.

2. crates/deadsync-shell/src/window.rs:95 (from 146d8241)

attributes = attributes.with_disallow_hidpi(!config.high_dpi);

error[E0599]: no method named `with_disallow_hidpi` found for struct `WindowAttributes`

The method does exist in winit 0.30.13; its trait is not in scope here. WindowAttributesExtMacOS is imported in window_size.rs, where this call used to live, but the call moved into window.rs during 146d8241 without its import. #618 imports the trait, cfg-gated to match the call site, and drops it from window_size.rs, which no longer uses it.

Reproduced on main with a clean worktree, stable rustc 1.96.0, aarch64-apple-darwin.

_{Edited: an earlier revision of this comment misdiagnosed (2) as the method being absent from winit 0.30.13. It is present; the trait was simply not imported.}