[GPU] Fix CI: always verify DMA ops are lowered, not just in debug builds.

* Move the unlowered-DMA verification walk out of #ifndef NDEBUG so it
  runs in both debug and release builds.
* Use notifyMatchFailure (silent) in the pattern guard instead of
  emitOpError, letting the post-pattern walk produce the single error.
* Remove the extra expected-error that only fired in debug builds.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: lialan

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

@@ -294,9 +294,9 @@ struct LowerCoalescedGatherDMAPattern final
       if (!hasAMDGPUFatRawBufferAddressSpace(srcType)) {
         for (Attribute attr : *inBounds) {
           if (!cast<BoolAttr>(attr).getValue()) {
-            dmaOp.emitOpError("in_bounds with OOB dimensions requires "
-                              "fat_raw_buffer address space on source");
-            return failure();
+            return rewriter.notifyMatchFailure(
+                dmaOp, "in_bounds with OOB dimensions requires "
+                       "fat_raw_buffer address space on source");
           }
         }
       }
@@ -503,7 +503,6 @@ struct AMDGPULowerCoalescedDMAToGatherLDSPass final
 
     walkAndApplyPatterns(funcOp, std::move(patterns));
 
-#ifndef NDEBUG
     // Verify all CoalescedGatherDMAOps were lowered. Currently, we require all
     // ops to be successfully lowered. In the future, a fallback lowering path
     // (e.g., using global_load) could handle ops that don't match the pattern.
@@ -514,7 +513,6 @@ struct AMDGPULowerCoalescedDMAToGatherLDSPass final
     if (result.wasInterrupted()) {
       return signalPassFailure();
     }
-#endif // NDEBUG
   }
 };
 } // namespace

compiler/src/iree/compiler/Codegen/Common/GPU/test/amdgpu_lower_coalesced_dma_to_gather_lds.mlir

@@ -1291,7 +1291,6 @@ func.func @no_lower_oob_without_fat_raw_buffer(
     attributes {hal.executable.target = #executable_target_rocm_hsaco_fb,
                 translation_info = #translation_64} {
   scf.forall (%arg6) in (64) {
-    // expected-error @+2 {{in_bounds with OOB dimensions requires fat_raw_buffer address space on source}}
     // expected-error @+1 {{failed to lower coalesced_gather_dma op}}
     iree_gpu.coalesced_gather_dma %source into %dest lane(%arg6) in_bounds [false, true] :
       memref<2x128xf32>,