block: shrink struct bio down to 2 cache lines again

Commit bcf2843b3f8f added ->bi_error to cleanup the error passing
for struct bio, but that ended up adding 4 bytes and a 4 byte hole
to the size of struct bio. For a clean config, that bumped it from
128 bytes, to 136 bytes, on x86-64.

The ->bi_flags member is currently an unsigned long, but it fits
easily within an int. Change it to an unsigned int, adjust the
the pool offset code, and move ->bi_error into the new hole. Then
we end up with a 128 byte bio again.

Change the bio flag set/clear to use cmpxchg to ensure we don't
lose any flags when manipulating them.

Signed-off-by: Jens Axboe <axboe@fb.com>