[Backend] Fix arith.index_cast lowering to LLVM. (#4704)

Fix crash when lowering `arith.index_cast` to llvm when the cast
truncates or extends the i32 index.

`createDestOps` should not `replaceOpWithNewOp` but simply `createOp`
because the call-site (`matchAndRewrite`) calls `replaceOp` later on.

lib/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVM.cpp

@@ -610,10 +610,9 @@ struct IndexCastOpLowering
     if (targetBits == sourceBits)
       return {operands[0][0]};
     if (targetBits < sourceBits)
-      return {rewriter.replaceOpWithNewOp<LLVM::TruncOp>(op, elemTy,
-                                                         operands[0][0])};
-    return {
-        rewriter.replaceOpWithNewOp<LLVM::SExtOp>(op, elemTy, operands[0][0])};
+      return {
+          rewriter.create<LLVM::TruncOp>(op.getLoc(), elemTy, operands[0][0])};
+    return {rewriter.create<LLVM::SExtOp>(op.getLoc(), elemTy, operands[0][0])};
   }
 };