xfrm: move policy_bydst RCU sync from per-netns .exit to .pre_exit

xfrm: move policy_bydst RCU sync from per-netns .exit to .pre_exit

The struct pernet_operations docstring in include/net/net_namespace.h
explicitly warns against blocking RCU primitives in .exit handlers:

    Exit methods using blocking RCU primitives, such as
    synchronize_rcu(), should be implemented via exit_batch.
    [...]
    Please, avoid synchronize_rcu() at all, where it's possible.

    Note that a combination of pre_exit() and exit() can
    be used, since a synchronize_rcu() is guaranteed between
    the calls.

xfrm_policy_fini() violates this: it calls synchronize_rcu() before
freeing the policy_bydst hash tables (so no RCU reader is mid-
traversal at free time), but runs from xfrm_net_ops.exit -- once per
namespace -- so a cleanup_net() of N namespaces pays N full RCU
grace periods serially.

Use the documented pre_exit/exit split. Move the policy flush (and
the workqueue drains it depends on) into a new .pre_exit handler;
xfrm_policy_fini() then runs in .exit and frees the hash tables
after the synchronize_rcu_expedited() that cleanup_net() guarantees
between the two phases. Providing O(1) RCU grace periods per batch
instead of O(N).

Observed on Linux 6.18 with a workload doing unshare(CLONE_NEWNET)
at ~13/sec sustained: cleanup_net() and the netns_wq rescuer kthread
both stuck in xfrm_policy_fini()'s synchronize_rcu(), >300k struct
net accumulated in the cleanup queue, Percpu in /proc/meminfo climbed
to 130+ GB on 256-CPU hosts, and memcg OOMs followed. setup_net and
__put_net counts were balanced, ruling out a refcount leak.

Fixes: 069daad4f2ae ("xfrm: Wait for RCU readers during policy netns exit")
Signed-off-by: Usama Arif <usama.arif@linux.dev>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>