Optimize rollback sync-test performance (#86)

## Summary

- remove immutable stage geometry and derived animation/collision data from rollback snapshots
- avoid a duplicate full skeletal hierarchy traversal during root-motion sampling
- cache root-motion cycle deltas per animation clip
- optimize the game crate in development builds
- make Tracy profiling opt-in via the `tracy` feature

Reviewed-on: #86
Co-authored-by: Álex Román Núñez <eirexe@eirteam.moe>

Cargo.toml

@@ -9,6 +9,7 @@ authors = [""]
 [features]
 default = ["development", "debug_features", "gamecube-adapter"]
 development = ["bevy/dynamic_linking", "bevy/file_watcher", "bevy/embedded_watcher", "hitbox-viewer"]
+tracy = ["bevy/trace_tracy"]
 debug_features = ["bevy/free_camera"]
 gamecube-adapter = ["dep:gc-adapter-neo"]
 hitbox-viewer = ["bevy/bevy_dev_tools", "dep:egui_extras"]
@@ -44,6 +45,11 @@ fixed_math = "0.4.1"
 [profile.dev.package."*"]
 opt-level = 3
 
+# The rollback simulation runs multiple times per rendered frame in sync-test
+# mode, so optimizing the game crate is important even for development builds.
+[profile.dev]
+opt-level = 1
+
 [profile.production]
 inherits = "release"
 lto = "thin"
@@ -56,4 +62,4 @@ gc-adapter-neo = { version = "1.1.1", features = ["libusb"], optional = true }
 [[bin]]
 name = "hitbox-editor"
 path = "src/bin/hitbox-editor.rs"
-required-features = ["hitbox-viewer"]
+required-features = ["hitbox-viewer"]