It's better organized now, I think - nodes only offer a single main way of getting the fundamental scalar (as a string), and now we can specialize a single template to read specific types.
In particular:
1. Windows CR/LF weren't read properly (issue #11)
2. Scanning wasn't reading EOF properly
3. Documents may be empty (this was old, I think)
Also fixed some VS2008 warnings on /W4.
We were getting memory leaks (as told by the CRT detectors, which I also added), and there's really no reason (as long as we're careful) to use pointers there.
