[LLVMGPU] Make FP8 VMFMA intrinsic discoverable by KernelConfig (iree-org#19022)

Making VMFMA_F32_16x16x32_F8E4M3FNUZ visible/accessible from
`getVirtualIntrinsics` S.T we can eventually pick this intrinsic from
KernelConfig.

Right now since by default since we always choose the "first" found
intrinsic, it is practically guaranteed to never pick the virtual
intrinsics by heuristics. Hence we are only specifying the intrinsics
through tuning/spec script.

However since we are planning to add some heuristics in the future on
KernelConfig to be able to select this virtual intrinsic when necessary,
we are making it visible now.

Signed-off-by: Stanley Winata <[email protected]>

Author: raikonenfnu

compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp

@@ -655,6 +655,8 @@ SmallVector<MMAIntrinsic> MMAAttr::getVirtualIntrinsics() const {
     return {MMAIntrinsic::VMFMA_F32_16x16x32_F16};
   case MMAIntrinsic::MFMA_F32_32x32x8_F16:
     return {MMAIntrinsic::VMFMA_F32_32x32x16_F16};
+  case MMAIntrinsic::MFMA_F32_16x16x32_F8E4M3FNUZ:
+    return {MMAIntrinsic::VMFMA_F32_16x16x32_F8E4M3FNUZ};
   default:
     return {};
   }