[LLVMGPU] Add multi-row vector reduction configuration (#73)

This is to speed up matvec. The new configuration is experimental and
only applied on ROCm targets.

compiler/src/iree/compiler/Codegen/Common/GPU/VectorReductionToGPU.cpp

@@ -196,7 +196,7 @@ class VectorReductionToGPUPass
       bool expandSubgroupReduction,
       std::function<int(func::FuncOp)> getWarpSize)
       : expandSubgroupReduction(expandSubgroupReduction),
-        getWarpSize(getWarpSize) {}
+        getWarpSize(std::move(getWarpSize)) {}
 
   void getDependentDialects(DialectRegistry &registry) const override {
     registry.insert<scf::SCFDialect, memref::MemRefDialect, gpu::GPUDialect,