Use AtomicF32 for the smoothing

This should at least not have a big performance impact since we can use
relaxed memory ordering here. AtomicCell always uses acquire/release
ordering.

Cargo.toml

@@ -22,6 +22,7 @@ members = [
 [dependencies]
 nih_plug_derive = { path = "nih_plug_derive" }
 
+atomic_float = "0.1"
 cfg-if = "1.0"
 crossbeam = "0.8"
 lazy_static = "1.4"