Improve strings provider shrinking order

[DRMacIver]

This is a piece of work I punted on doing for the initial strings implementation in #23 - currently we will shrink towards lower codepoints, which is sensible enough but doesn't do a very good job for the very low codepoint ascii order. In hypothesis-python we shrink towards ascii '0' (i.e. codepoint 48). We should do that here too.