This just reorganizes some ghastly goto-and-retry-spaghetti in the core of
the ext3 block allocator.

I wrote this ages ago in preparation for fixing the find_next_usable_block()
CPU pigginess problem, but that proved to be quite nontrivial.

The patch has been in -mm for a long time and Martin has recently confirmed
that it introduces no performance regression in SDET and kernbench.