ublk: implement NUMA-aware memory allocation

Implement NUMA-friendly memory allocation for ublk driver to improve
performance on multi-socket systems.

This commit includes the following changes:

1. Rename __queues to queues, dropping the __ prefix since the field is
now accessed directly throughout the codebase rather than only through
the ublk_get_queue() helper.

2. Remove the queue_size field from struct ublk_device as it is no longer
needed.

3. Move queue allocation and deallocation into ublk_init_queue() and
ublk_deinit_queue() respectively, improving encapsulation. This
simplifies ublk_init_queues() and ublk_deinit_queues() to just
iterate and call the per-queue functions.

4. Add ublk_get_queue_numa_node() helper function to determine the
appropriate NUMA node for a queue by finding the first CPU mapped
to that queue via tag_set.map[HCTX_TYPE_DEFAULT].mq_map[] and
converting it to a NUMA node using cpu_to_node(). This function is
called internally by ublk_init_queue() to determine the allocation
node.

5. Allocate each queue structure on its local NUMA node using
kvzalloc_node() in ublk_init_queue().

6. Allocate the I/O command buffer on the same NUMA node using
alloc_pages_node().

This reduces memory access latency on multi-socket NUMA systems by
ensuring each queue's data structures are local to the CPUs that
access them.

Reviewed-by: Caleb Sander Mateos <csander@purestorage.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>