[TritonGPU] Run Gluon canonicalizer in LLVM lowering when ConSan is on (#9396)

Regular canonicalizer will rewrite layout conversions

Author: Mogball

third_party/nvidia/backend/compiler.py

@@ -364,7 +364,7 @@ def make_llir(self, src, metadata, options, capability):
         if "consan" in options.instrumentation_mode:
             # Call ConcurrencySanitizerPass here, before allocating global scratch memory but after allocating tensor and shared
             passes.ttgpuir.add_concurrency_sanitizer(pm)
-            passes.common.add_canonicalizer(pm)
+            passes.gluon.add_canonicalizer(pm)
             passes.common.add_cse(pm)
         passes.ttgpuir.add_allocate_global_scratch_memory(pm)
         nvidia.passes.ttnvgpuir.add_proxy_fence_insertion(pm, capability)