[phobos] Definition of Random Access Range

[David Simcha]

Should we just redefine random access ranges such that they must either 
have a length or be infinite?  I'm trying to improve the unit tests in 
std.range in preparation for some bug fixing and consistency 
improvements and I'm finding the mother lode of bugs in the process 
(which I'll file or fix relatively soon).  I basically made a dummy 
range template, made a tuple of instantiations with every combination of 
(ref, non-ref returns), (length, no length) and (input, forward, 
bidirectional, random access) and tried to instantiate the higher order 
ranges with every one of these that made sense.  One recurring theme is 
that the case of a random access range that isn't infinite and doesn't 
have a length is not properly dealt with.

Since I can't imagine how a range could offer O(1) random access w/o 
either knowing its length or being infinite, I suggest we just add this 
requirement to the definition of a random access range and be done with 
it.  Counterexamples?