This makes NestingLevel more properly model the global nesting level of a
token, instead of the nesting level local to the current annotated line.
The NestingLevel field can then be used to determine the start and end of
token alignment sequences more precisely.