This momentarily breaks the iterator fault injection test, but I would
have had to squash too many commits to make it work right, so it
appeared to be the better option to disable the respective test for some
commits.

this is to make reading disk objects more efficient later on.

This patch triggered c93.vtc failing, so we add a bit of a hack to avoid
a problem which might need more attention later:

FCO_MAX_REGIONS is an issue for chunked encoding objects (with growing
size). We have not yet implemented the best strategy and just tried
to always allocate the largest possible seglist to not use up more
regions than necessary, but small memory configurations do not support
the maximum seglist (4 MB).

previously, we just sized the disk seglist to fit.

Now we chose the size such that the memory seglist corresponding to the
disk seglist fits a power of two page.

This allows us to shrink the fellow_cache_obj allocation for the
fellow_obj_get (from disk) case from 8KB to 512 bytes.

The root cause for the massively oversized fco allocation was that
the nseg_guess heuristic could not take into account the wsl (size of
the actual object), so it had to assume that all of the disk object's
size was taken up for segments in the disk object embedded seglist.