Where possible, switch from early bio-based IO accounting (at the time
DM clones each incoming bio) to late IO accounting just before each
remapped bio is issued to underlying device via submit_bio_noacct().

Allows more precise bio-based IO accounting for DM targets that use
their own workqueues to perform additional processing of each bio in
conjunction with their DM_MAPIO_SUBMITTED return from their map
function. When a target is updated to use dm_submit_bio_remap() they
must also set ti->accounts_remapped_io to true.

Use xchg() in start_io_acct(), as suggested by Mikulas, to ensure each
IO is only started once.  The xchg race only happens if
__send_duplicate_bios() sends multiple bios -- that case is reflected
via tio->is_duplicate_bio.  Given the niche nature of this race, it is
best to avoid any xchg performance penalty for normal IO.

For IO that was never submitted with dm_bio_submit_remap(), but the
target completes the clone with bio_endio, accounting is started then
ended and pending_io counter decremented.
Reviewed-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>