Renumber slot indexes uniformly instead of spacing according to the number of defs.

There are probably much larger speedups to be had by renumbering locally instead
of looping over the whole function. For now, the greedy register allocator is
25% faster.

llvm-svn: 126926