The tokenizer caches output for common chunks, for efficiency. This
cache is be invalidated when the tokenizer rules change, e.g. when a new
special-case rule is introduced. That's what was causing #1061.
When the cache is flushed, we free the intermediate token chunks.
I *think* this is safe --- but if we start getting segfaults, this patch
is to blame. The resolution would be to simply not free those bits of
memory. They'll be freed when the tokenizer exits anyway.
