lib/bdev: submit child splitted I/Os with public memory domain

If a bdev doesn't support memory domain, a bounce-buffer is allocated
and data is either pulled to the bounce-buffer (WRITEs) or a push from
the bounce-buffer is scheduled after completion (READs). If this
fallback is used, `bdev_io->u.bdev.memory_domain` is reset to NULL.

The fallback is executed before splitting. Thus, the splitted I/Os
should not use the memory domain of the parent I/O if the fallback has
already been executed (data has been pulled or a push was scheduled
after the completion of the parent I/Os).

The current implementation causes a wrong pull/push request between the
parent I/O's bounce-buffer and the child I/O's bounce-buffer, which is
nonsensical because the parent I/O's bouce-buffer doesn't belong to the
memory domain region.

I assume that:
- `bdev_io->internal.memory_domain` is supposed to be the memory domain
  at submission, to be used for push/pull on bounce-buffers
- `bdev_io->u.bdev.memory_domain` is supposed to be the memory domain
  used (e.g. by modules) to submit child I/Os
The Bdev layer breaks this assuption when it submits splitted I/Os. I
suggest to use `bdev_io->u.bdev.memory_domain` when submitted splitted
I/Os.

Reproducer:
- create a bdev with an optimal I/O boundary
- submit an I/O that spans multiple I/O boundaries with a memory domain
- check the requested transfers

Change-Id: I03cc76ade16500e74d547b9f07f3a74f5e5e391b
Signed-off-by: Nathan Claudel <nclaudel@kalrayinc.com>
Reviewed-on: https://review.spdk.io/c/spdk/spdk/+/25739


Tested-by: SPDK Automated Test System <spdkbot@gmail.com>
Reviewed-by: Daniil Ushkov <daniil.ushkov@yandex.ru>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>