Track consumed token count in ParserState (and ParseError)

[phadej]

Relying on (user settable) SourcePos for deciding which parse is longer (in mergeError) is unreliable.

FWIW megaparsec tracks the count.

The #96 is different, as it adds character count to SourcePos. Yet, both changes are breaking ones, as ErrorPos "constructors" are part of public interface.