spirv_hip_fp16.h header file updates

[jjennychen]

Partially updated the spirv_hip_fp16.h based on rocm release 6.1.0 and added the needed ocml functions to spirv_math_fwd.h (tested the ocml functions and worked well).
The parts of spirv_hip_fp16.h left unmodified includes:

• Updates in macros defined in other header files
  • Left unchanged as it may impact multiple other header files
  • For example, #if __HIP_CLANG_ONLY__ was updated to #if defined(__clang__) && defined(__HIP__) in the new release
• #includes for AMD-specific header files
  • Did not include the header files such as <hip/amd_detail/amd_hip_common.h> and "hip/amd_detail/host_defines.h" (it seems like we do have our host_defines.h in chipStar, but it seems to still be in progress)
• Newly added shuffle functions
  • Did not add the shuffle functions from new header in as they are (to my knowledge) hardware-dependent and may need our own implementation
  • The shuffle functions are the following:

        __device__
        inline
        __half __shfl(__half var, int src_lane, int width = warpSize) {
           union { int i; __half h; } tmp; tmp.h = var;
           tmp.i = __shfl(tmp.i, src_lane, width);
           return tmp.h;
        }
        __device__
        inline
        __half2 __shfl(__half2 var, int src_lane, int width = warpSize) {
           union { int i; __half2 h; } tmp; tmp.h = var;
           tmp.i = __shfl(tmp.i, src_lane, width);
           return tmp.h;
        }
        __device__
        inline
        __half __shfl_up(__half var, unsigned int lane_delta, int width = warpSize) {
           union { int i; __half h; } tmp; tmp.h = var;
           tmp.i = __shfl_up(tmp.i, lane_delta, width);
           return tmp.h;
        }
        __device__
        inline
         __half2 __shfl_up(__half2 var, unsigned int lane_delta, int width = warpSize) {
            union { int i; __half2 h; } tmp; tmp.h = var;
            tmp.i = __shfl_up(tmp.i, lane_delta, width);
            return tmp.h;
         }
         __device__
         inline
         __half __shfl_down(__half var, unsigned int lane_delta, int width = warpSize) {
            union { int i; __half h; } tmp; tmp.h = var;
            tmp.i = __shfl_down(tmp.i, lane_delta, width);
            return tmp.h;
         }
         __device__
         inline
         __half2 __shfl_down(__half2 var, unsigned int lane_delta, int width = warpSize) {
            union { int i; __half2 h; } tmp; tmp.h = var;
            tmp.i = __shfl_down(tmp.i, lane_delta, width);
            return tmp.h;
         }
         __device__
         inline
         __half __shfl_xor(__half var,  int lane_mask, int width = warpSize) {
            union { int i; __half h; } tmp; tmp.h = var;
            tmp.i = __shfl_xor(tmp.i, lane_mask, width);
            return tmp.h;
         }
         __device__
         inline
          __half2 __shfl_xor(__half2 var,  int lane_mask, int width = warpSize) {
             union { int i; __half2 h; } tmp; tmp.h = var;
             tmp.i = __shfl_xor(tmp.i, lane_mask, width);
             return tmp.h;
         }

(Note: __llvm_rcp_f16 and __llvm_rcp_2f16 defines are removed as they are no longer used in the updated header file (implementation was changed from using those to using direct operations (such as division))