We're currently using a libxml buffer, which requires copying the buffer
to zend_strings every time we want to output the string. Furthermore,
its use of the system allocator instead of ZendMM makes it not count
towards the memory_limit and hinders performance.
This patch adds a custom writer such that the strings are written to a
smart_str instance, using ZendMM for improved performance, and giving
the ability to not copy the string in the common case where flush has
empty set to true.
Closes GH-16120.
