Optimize the performance of serialize method

[TennyZhuang]

Concisely describe the proposed feature

Currently, we use virtual std::string serialize() = 0; in Expression interface. The serialize which returns a string is concise, but it's inefficient for a tree-style data structure. e.g. If our tree has 10 level and 30 nodes, we have to do at least 30 buffer allocations.

Describe the solution you'd like (if any)

Maybe we can pass a std::stringstream in the serialize, or some other custom appended, e.g.:

void virtual serialize(std::stringstream&);

We can use less allocations to improve the performance, e.g., for GlobalPtrExpression:

void GlobalPtrExpression::serialize(std::stringstream& builder) {
  builder << snode ? snode->get_node_type_name_hinted() : var.serialize();
  builder << '[';
  for (int i = 0; i < (int)indices.size(); i++) {
    indices.exprs[i]->serialize(builder);
    if (i + 1 < (int)indices.size())
      builder << ', ';
  }
  builder << ']';
}

Additional comments

[skyzh]

std::stringstream should be std::stringstream&?

[k-ye]

Good point! Feel free to send a patch if you'd like to :-)

[strongoier]

Fixed by @TennyZhuang in #3221.