string `length` performance trap

[JeffBezanson]

I recently saw some code that used while i <= length(s) where s is a String. This can happen e.g. when porting code that assumes all data is ASCII. Obviously one answer is "don't do that", but it would be nice if we could handle this better. Two ideas:

• Tell LLVM that length(::String) is pure and can be hoisted out of loops.
• Use the top bit of the length field to store whether a string is ASCII, speeding up length and isascii in that case.

[Keno]

For the top bit, last time we discussed this, I had suggested that length and isascii could automatically set that if applicable the first time they're run. Of course we'd need to update things like egality and it'd make String even more weird of an object type, but probably worthwile. Of course telling LLVM it can LICM it is good in either case.