[mlir][sparse] fix memory leakage in concatenate rewriter.

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorRewriting.cpp

for (Value input : op.getInputs()) {
       auto d = input.getType().cast<RankedTensorType>().getShape()[conDim];
       assert(!ShapedType::isDynamic(d));
       offset = rewriter.create<arith::AddIOp>(loc, offset,
                                               constantIndex(rewriter, loc, d));
       cooBuffer = foreachOp.getResult(0);
     }
 
     cooBuffer = rewriter.create<LoadOp>(loc, cooBuffer, true);
-    rewriter.replaceOpWithNewOp<ConvertOp>(op, rtp, cooBuffer);
+    Value converted =
+        rewriter.create<ConvertOp>(loc, rtp, cooBuffer).getResult();
+    rewriter.create<DeallocTensorOp>(loc, cooBuffer);
+    rewriter.replaceOp(op, converted);
     return success();
   }
 };
 
 /// Sparse rewriting rule for the convert operator.
 struct ConvertRewriter : public OpRewritePattern<ConvertOp> {
   using OpRewritePattern::OpRewritePattern;
   LogicalResult matchAndRewrite(ConvertOp op,