Save one canonicalization pass in TOSA input conversion (#8454)

This saves one run of canonicalization and it also fixes a performance
regression for PoseNet on Mali GPUs.

iree/compiler/InputConversion/TOSA/Passes.cpp

diff --git a/iree/compiler/InputConversion/TOSA/Passes.cpp b/iree/compiler/InputConversion/TOSA/Passes.cpp
index f4b86a9..0b3bf54 100644
--- a/iree/compiler/InputConversion/TOSA/Passes.cpp
+++ b/iree/compiler/InputConversion/TOSA/Passes.cpp
@@ -48,8 +48,6 @@
   passManager.addNestedPass<FuncOp>(tosa::createTosaToStandard());
 
   passManager.addNestedPass<FuncOp>(IREE::Flow::createStripSignednessPass());
-  passManager.addNestedPass<FuncOp>(mlir::createCanonicalizerPass());
-
   passManager.addNestedPass<FuncOp>(createLinalgQuantizedMatmulToMatmulPass());
   passManager.addNestedPass<FuncOp>(mlir::createCanonicalizerPass());