Sync upstream build profiles through 710eb8f5

Apply upstream 710eb8f531 on the previously integrated c2fc3cf3 snapshot: disable dev ThinLTO and build dev/play dependencies as one codegen unit. Release and ship profiles, MoonWell integration and ruRU localization remain unchanged. Regenerated MAP has no delta.

Verified on Windows: cargo fmt --all -- --check and locked workspace metadata pass; the applied patch matches upstream. Full land gates run before merge. The upstream performance measurements were on macOS; no Windows speedup is claimed. No declared account for live login.

Cargo.toml

@@ -46,11 +46,19 @@ kira = { path = "third_party/kira" }
 # Dependencies at opt-level 3 in dev: at opt-level 0, avian/parry collision drops a run to <1 fps.
 [profile.dev.package."*"]
 opt-level = 3
+# One codegen unit each, since `lto = "off"` below drops the ThinLTO that inlines across a crate's
+# units: at 16 units the dependencies cost +1.2 cpu_ms at the leg pin, at one nothing, and a cold
+# build is no slower.
+codegen-units = 1
 
 # Our crates at opt-level 1: Bevy's generic hot paths monomorphize into them, and at opt-level 0
-# Stormwind drops from a pinned 60 fps to ~45; the warm incremental rebuild cost is ~0.
+# Stormwind drops from a pinned 60 fps to ~45.
 [profile.dev]
 opt-level = 1
+# No ThinLTO: above opt-level 0 rustc runs it over a crate's own codegen units by default, and an
+# incremental rebuild redoes it whole: 19 s of a one-line `benilla-app` edit's 25 s, and half its
+# 5.3 GB. Off, our crates cost +0.7 cpu_ms (+6 %) at the leg pin.
+lto = "off"
 
 # No debuginfo trimming: `debug = "line-tables-only"` measures no recurring win on macOS + sccache.
 
@@ -71,3 +79,8 @@ incremental = true
 # Not for speed: under `incremental`, thin-local LTO leaves promoted `anon.*.llvm.*` symbols
 # unresolved and the link fails intermittently (rustc #86049).
 lto = "off"
+
+# One codegen unit each, as in dev: at 16 the dependencies cost +1.2 cpu_ms against `release` at
+# the leg pin, at one the profile is +0.5, and a cold build is no slower.
+[profile.play.package."*"]
+codegen-units = 1