The attribute name is valid until the next attribute or next node.
This simplifies the code a bit: fewer error paths, and free() calls.

A few potential memory leaks were fixed too.