Add option to cache node names across transformers

[stanhu]

In 4.6.2, caa558a significantly cut down on the memory usage by downcasing the node name once.

However, someone reported in #177 reported that this broke certain transformers that modified the node name since the name was cached.

I'd suggest that it would still be nice to have a way to support that optimization for applications that don't change the node name. In my sample test, I see that 11 MB of RAM is being allocated, and with thousands of these documents being processed every second on GitLab, we are seeing slowdowns due to garbage collection.

Thoughts?

[rgrove]

Seems like a simple solution might be to reimplement the change from caa558a but compare node_name with node.name after each iteration and update it if the node name changes.