1) Remember the granule position of the last page we've seen from
    the current link and save the first page of the next link as
    long as we're scanning forward.
   This knocks almost 10% off the number of seeks for large links.
   For smaller links the improvement is much larger.
2) Only use pairs of close-by serial numbers to estimate link
    start locations (assuming they're above our start threshold).
   This gives a minor (<2%) improvement, which might be in the
    noise, but as it doesn't appear to hurt and is faster, might as
    well.
3) Eliminate a redundant check in op_pcm_seek_page_impl().