[Codegen][GPU] Move PackToIntrinsics after workgroup tiling (iree-org…

…#17950)

Doing `PackToIntrinsics` before workgroup tiling is putting lower level
lowering details too early in the pipeline, especially given that
padding needs to happen at the workgroup level, not before that.

There is nothing to test for here, the generated IR remains the same.

compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp

@@ -289,13 +289,12 @@ void addGPUVectorizationPassPipeline(OpPassManager &funcPassManager) {
 //===---------------------------------------------------------------------===//
 
 void addGPUTileAndFusePassPipeline(OpPassManager &funcPassManager) {
+  tileAndDistributeToWorkgroup(funcPassManager);
 
   // Step 1. Promote matmul operands and pack to intrinsic shapes.
   funcPassManager.addPass(createGPUPromoteMatmulOperandsPass());
   funcPassManager.addPass(IREE::GPU::createPackToIntrinsicsPass());
 
-  tileAndDistributeToWorkgroup(funcPassManager);
-
   // Step 2. Tile and fuse tileable ops to reduction loops.
   {
     GPUApplyTilingLevelPassOptions options;