Overload unary - operator on at::vec::Vectorized to call neg() (pytorch#150568)

Makes Vectorized look even more like a scalar type, getting me closer to being able to use the same generic code with scalars and Vectorized (e.g., for sigmoid, which needs `exp(-x)`).

Pull Request resolved: pytorch#150568
Approved by: https://github.com/Skylion007
ghstack dependencies: pytorch#150380

Author: swolchok

aten/src/ATen/cpu/vec/vec_base.h

@@ -740,6 +740,11 @@ struct Vectorized {
   }
 };
 
+template <class T>
+Vectorized<T> inline operator-(const Vectorized<T>& a) {
+  return a.neg();
+}
+
 // There is an implicit conversion that would make this work if
 // these operators weren't template functions, but they are template
 // functions (and can't be moved to be non-member friends defined in

aten/src/ATen/test/vec_test_all_types.cpp

@@ -192,6 +192,11 @@ namespace {
             [](vec v) { return v.neg(); },
             createDefaultUnaryTestCase<vec>(TestSeed()),
             RESOLVE_OVERLOAD(filter_int_minimum));
+        test_unary<vec>(
+            NAME_INFO(negate), std::negate<ValueType<vec>>(),
+            [](vec v) { return -v; },
+            createDefaultUnaryTestCase<vec>(TestSeed()),
+            RESOLVE_OVERLOAD(filter_int_minimum));
     }
     TYPED_TEST(SignManipulationHalfPrecision, AbsNegate) {
       typedef enum  {