Reduce a little overhead (#871)

* some small overhead improvements

* use result_type in rms_norm

* remove release force

* fix + use non-vector version

* revert compile change

* fix ops

* a little more overhead

* a little more cleanup and overhead

mlx/array.h

@@ -392,6 +392,10 @@ class array {
         Dtype dtype,
         std::shared_ptr<Primitive> primitive,
         std::vector<array> inputs);
+
+   private:
+    // Initialize size, strides, and other metadata
+    void init();
   };
 
   // The ArrayDesc contains the details of the materialized array including the