[mlir][sparse] fix a bug in unpack op that used wrong compare predicate.

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

 /// length is larger, or creates a subview if it is smaller).
 static Value reallocOrSubView(OpBuilder &builder, Location loc, int64_t len,
                               Value buffer) {
   MemRefType memTp = getMemRefType(buffer);
   auto retTp = MemRefType::get(ArrayRef{len}, memTp.getElementType());
 
   Value targetLen = constantIndex(builder, loc, len);
   Value bufferLen = linalg::createOrFoldDimOp(builder, loc, buffer, 0);
-  Value reallocP = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ult,
+  // Reallocates if target length is greater than the actual buffer len.
+  Value reallocP = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ugt,
                                                  targetLen, bufferLen);
   scf::IfOp ifOp = builder.create<scf::IfOp>(loc, retTp, reallocP, true);
   // If targetLen > bufferLen, reallocate to get enough sparse to return.
   builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
   Value reallocBuf = builder.create<memref::ReallocOp>(loc, retTp, buffer);
   builder.create<scf::YieldOp>(loc, reallocBuf);
   // Else, return a subview to fit the size.
   builder.setInsertionPointToStart(&ifOp.getElseRegion().front());

mlir/test/Dialect/SparseTensor/sparse_pack.mlir

 // CHECK-LABEL:   func.func @sparse_unpack(
 // CHECK-SAME:      %[[VAL_0:.*]]: memref<?xindex>,
 // CHECK-SAME:      %[[VAL_1:.*]]: memref<?xi32>,
 // CHECK-SAME:      %[[VAL_2:.*]]: memref<?xf64>,
 // CHECK-SAME:      %[[VAL_3:.*]]: !sparse_tensor.storage_specifier
 // CHECK:           %[[VAL_4:.*]] = arith.constant 6 : index
 // CHECK:           %[[VAL_5:.*]] = arith.constant 0 : index
 // CHECK:           %[[VAL_6:.*]] = memref.dim %[[VAL_2]], %[[VAL_5]] : memref<?xf64>
-// CHECK:           %[[VAL_7:.*]] = arith.cmpi ult, %[[VAL_4]], %[[VAL_6]] : index
+// CHECK:           %[[VAL_7:.*]] = arith.cmpi ugt, %[[VAL_4]], %[[VAL_6]] : index
 // CHECK:           %[[VAL_8:.*]] = scf.if %[[VAL_7]] -> (memref<6xf64>) {
 // CHECK:             %[[VAL_9:.*]] = memref.realloc %[[VAL_2]] : memref<?xf64> to memref<6xf64>
 // CHECK:             scf.yield %[[VAL_9]] : memref<6xf64>
 // CHECK:           } else {
 // CHECK:             %[[VAL_10:.*]] = memref.subview %[[VAL_2]][0] [6] [1] : memref<?xf64> to memref<6xf64>
 // CHECK:             scf.yield %[[VAL_10]] : memref<6xf64>
 // CHECK:           }
 // CHECK:           %[[VAL_11:.*]] = arith.constant 12 : index
 // CHECK:           %[[VAL_12:.*]] = memref.dim %[[VAL_1]], %[[VAL_5]] : memref<?xi32>
-// CHECK:           %[[VAL_13:.*]] = arith.cmpi ult, %[[VAL_11]], %[[VAL_12]] : index
+// CHECK:           %[[VAL_13:.*]] = arith.cmpi ugt, %[[VAL_11]], %[[VAL_12]] : index
 // CHECK:           %[[VAL_14:.*]] = scf.if %[[VAL_13]] -> (memref<12xi32>) {
 // CHECK:             %[[VAL_15:.*]] = memref.realloc %[[VAL_1]] : memref<?xi32> to memref<12xi32>
 // CHECK:             scf.yield %[[VAL_15]] : memref<12xi32>
 // CHECK:           } else {
 // CHECK:             %[[VAL_16:.*]] = memref.subview %[[VAL_1]][0] [12] [1] : memref<?xi32> to memref<12xi32>
 // CHECK:             scf.yield %[[VAL_16]] : memref<12xi32>
 // CHECK:           }
 // CHECK:           %[[VAL_17:.*]] = memref.expand_shape %[[VAL_18:.*]] {{\[\[}}0, 1]] : memref<12xi32> into memref<6x2xi32>