Fix performance issue with nanstd and nanvar

Author: zhujun98

include/xtensor/xmath.hpp

@@ -2610,8 +2610,8 @@ namespace detail {
               XTL_REQUIRES(is_reducer_options<EVS>)>
     inline auto nanvar(E&& e, EVS es = EVS())
     {
-        decltype(auto) sc = detail::shared_forward<E>(e);
-        return nanmean<T>(square(sc - nanmean<T>(sc)), es);
+        auto cached_mean = nanmean<T>(e, es)();
+        return nanmean<T>(square(std::forward<E>(e) - cached_mean), es);
     }
 
     template <class T = void, class E, class EVS = DEFAULT_STRATEGY_REDUCERS,
@@ -2648,7 +2648,7 @@ namespace detail {
         // note: forcing copy of first axes argument -- is there a better solution?
         auto axes_copy = axes;
         using result_type = typename std::conditional_t<std::is_same<T, void>::value, double, T>;
-        auto inner_mean = nanmean<result_type>(sc, std::move(axes_copy));
+        auto inner_mean = eval(nanmean<result_type>(sc, std::move(axes_copy), evaluation_strategy::immediate));
 
         // fake keep_dims = 1
         auto keep_dim_shape = e.shape();