More backreference performance improvements [take 2]

[Bonzini]

> Since I'm now working on fixing the bad algorithm which causes the bad
> performance, and the fix might be going to spoil your optimizations
> based on cache profiling.

My fix is not based on cache profiling.  I did do cache profiling and it
turned out to be very good, so I exclusively concentrated on instruction
counts (the Ir column in the cachegrind output gives the number of L1 cache
hits for the instruction, which is practically the same as the instruction
count).  My 60% optimization does nothing more than limiting the lower and
upper bound of the inner loops in sift_states_bkref.

Paolo