No-one seems to have noticed, but there's a bug in our offline handling
which can cause the eh to loop forever when it tries to offline a device
(basically offline I/O rejections are done in prep, which doesn't work
for already prepped commands).

The attached fixes this, and also sweeps offline up into a fairly fully
fledged scsi state model.

With this applied I can now offline my root device under load without
causing a SCSI hang (ext2 BUG()s rather unhappily but at least that's
not a SCSI problem).