Rebase the PR #1520 to ROCm repo. (#1574)

* Implement hiprtc for codegen tests

* Introduce gemm_softmax_gemm to codegen.

* Fix codegen build issues.

* Address PR comments.

* Separate ck_host lib and gemm_softmax_gemm into different PR.

* Fix cmake.

* Replace ENV variable with CMake option for toggling hipRTC in codegen
tests.

* Address PR comments.

* fix clang format

* Add missing header in magic_division.hpp

* - Workaround for hipRTC content wrapper
- Move descriptor for gemm_softmax_gemm to different branch

* Fix formatting.

* Revert "Fix formatting."

This reverts commit b5209eaef4.

* formatting fix

* fixed header guard issues

* updated header guards

* updated data_type for new types

* fixed redefinition error

* Add codegen test for batched_gemm_softmax_gemm.

Signed-off-by: Mirza Halilcevic <mirza.halilcevic@amd.com>

* formatting fix

---------

Signed-off-by: Mirza Halilcevic <mirza.halilcevic@amd.com>
Co-authored-by: Dino Musić <dino.music@htecgroup.com>
Co-authored-by: Mirza Halilcevic <mirza.halilcevic@htecgroup.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: arai713 <67439843+arai713@users.noreply.github.com>
Co-authored-by: Astha Rai <astha.rai713@gmail.com>
Co-authored-by: Mirza Halilcevic <mirza.halilcevic@amd.com>

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -697,7 +697,7 @@ struct FastGelu
 
     template <typename Y, typename X>
     __device__ void operator()(Y& y, const X& x) const;
-#ifndef CK_CODE_GEN_RTC
+#if !defined(__HIPCC_RTC__) || !defined(CK_CODE_GEN_RTC)
     template <>
     __host__ void operator()<float, float>(float& y, const float& x) const
     {
@@ -709,7 +709,6 @@ struct FastGelu
         y               = x / (1.f + emu);
     }
 #endif
-
     // device code, use lower precision "__ocml_exp_f32" and "rcp"
     template <>
     __device__ void operator()<float, float>(float& y, const float& x) const