optimization: consider alternative document free mechanisms that will allow setting `RUBY_TYPED_FREE_IMMEDIATELY`

[flavorjones]

See 0804380 for more context, but essentially:

• we parent unparented nodes at the last minute so that xmlFreeDoc frees all the underlying memory for us
• if two text nodes happen to end up siblings during that process, they are merged and libxml2 may allocate additional memory via ruby_xmalloc
• it may be unsafe to allocate memory during GC, warnings will be raised

Working on #2807 made me realize that we're doing more work during the deallocate function than I thought -- so it's both slower and less safe than it could be.

I have at least one idea, which is instead of parenting the unparented nodes, just call xmlFreeNode on them to avoid the parenting overhead.

[byroot]

just call xmlFreeNode on them to avoid the parenting overhead.

I tried that and got some crashes because the pointers weren't allocated, not too shure what kind of shenanigans libxml does internally, but I suspect it might have some kind of arenas.