PR #8919: Turn the HLO verifier always on around float normalization on GPU

Imported from GitHub PR openxla/xla#8919

Currently, the GPU pipeline only turns the verifier around float normalization on only in debug. If float normalization generates ill-typed IR, the compiler can produce invalid code without anything failing. This patch turns the verification always on.
Copybara import of the project:

--
8606a6b140091a099c541e1d48f6f685828c58d8 by Jaroslav Sevcik <jsevcik@nvidia.com>:

Turn the HLO $verifier always on around float normalization on GPU

--
e3dc564b580f5fa3e0663a87326451d0fab3db1c by Jaroslav Sevcik <jsevcik@nvidia.com>:

Run verifier only once

--
13f4f25deb24250ed71775b514fad4de3071e2b8 by Jaroslav Sevcik <jsevcik@nvidia.com>:

Add a context string to the verifier pass

--
c97786e4024aa97062f0c48c117168d48360e318 by Jaroslav Sevcik <jsevcik@nvidia.com>:

Only verify in non-debug builds

Merging this change closes #8919

PiperOrigin-RevId: 605295966

Author: jaro-sevcik

third_party/xla/xla/service/gpu/gpu_compiler.cc

@@ -1399,6 +1399,21 @@ absl::Status GpuCompiler::OptimizeHloPostLayoutAssignment(
   pipeline.AddPass<HloCSE>(/*is_layout_sensitive=*/true,
                            /*only_fusion_computations=*/false,
                            /*ignore_control_dependencies=*/true);
+
+#ifdef NDEBUG
+  // Verify the module in non-debug builds. For debug builds, the verifier
+  // already runs after every pass.
+  pipeline.AddPass<HloVerifier>(
+      std::make_unique<DefaultVerifierMetadata>(
+          HloVerifierOpts{}
+              .MakeLayoutSensitive()
+              .WithInstructionCanChangeLayout(
+                  LayoutAssignment::InstructionCanChangeLayout)
+              .VerifyBroadcastDimensionsOrder()
+              .VerifyReshapeIsBitcast()),
+      "end-of-post-layout_assignment");
+#endif  // NDEBUG
+
   TF_RETURN_IF_ERROR(pipeline.Run(hlo_module).status());
 
   return absl::OkStatus();