Another round of optimizations

[kddnewton]

• Use rb_ary_store instead of rb_ary_push when the array is right-sized
• Put a cache for the constant pool on the parser since we re-hash the same bounds between the lexer and the parser (I actually would like to solve this differently, but for now it's like an 8% win)
• Lazily unpack the line offsets the first time they are accessed so if you are a Ruby API consumer that does not touch line numbers you don't pay for the array building
• Fast-path out of parse_arguments_list when a token cannot possibly start an argument.
• Use an arena for building the stack that builds the AST in Ruby