Change std::hash, memory layout, and default allocation sizes for
std::unordered_map. More advanced storage and probing techniques (as seen in
absl::node_hash_map) could result in microbenchmark improvements of 300-400%
over most unordered_map implementations.