Fix for AdaDelta (#603)

- state was being read from parameter "s"
- but being stored in parameter "u"

Author: davidkoski

python/mlx/optimizers.py

@@ -284,7 +284,7 @@ def apply_single(
         eps = self.eps
 
         v = state.get("v", mx.zeros_like(gradient))
-        u = state.get("s", mx.zeros_like(gradient))
+        u = state.get("u", mx.zeros_like(gradient))
 
         v = rho * v + (1 - rho) * mx.square(gradient)
         d = mx.sqrt(u + eps) / mx.sqrt(v + eps) * gradient