Previously, the lexer tried to keep only part of each source file in
memory, the part that wasn't fully parsed yet. With this commit, the
lexer holds the entire file in memory, even parts that are no longer
needed. This should make it easier to produce better error messages.