habanalabs: use GFP_ATOMIC under spin lock

A spin lock is taken here so we should use GFP_ATOMIC.

Fixes: 0feaf86d ("habanalabs: add virtual memory and MMU modules")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>