WMMA grouped conv fwd large tensor extra flavors (#3582)

* Additional flavors for WMMA conv fwd large tensor

- added F16/BF16 clamp operation
- added F16/BF16 bias_clamp operation
- small modification to the device code to accomodate extra tensors

* changed strategy to handle GemmArgs array

* Adding generic instance

* Added generic instance to clamp and bias_clamp ops

include/ck/utility/array.hpp

@@ -6,6 +6,8 @@
 
 #include "functional2.hpp"
 #include "sequence.hpp"
+#include <type_traits>
+#include <cassert>
 
 namespace ck {
 
@@ -27,6 +29,15 @@ struct Array
 
     __host__ __device__ constexpr TData& operator()(index_t i) { return At(i); }
 
+    template <typename... Args>
+    __host__ constexpr auto Emplace(index_t i, Args&&... args)
+        -> std::enable_if_t<std::is_nothrow_constructible_v<TData, Args&&...>>
+    {
+        assert(i >= 0 && i < NSize);
+        mData[i].~TData();
+        new(mData + i) TData(ck::forward<Args>(args)...);
+    }
+
     template <typename T>
     __host__ __device__ constexpr auto operator=(const T& a)
     {