net/sched: account classifier filter allocations to memcg

In the Linux kernel, the following vulnerability has been resolved:

net/sched: account classifier filter allocations to memcg

Allocations in the tc classifier *_change() paths (filter objects,
per-CPU counters, and per-filter aux data) use plain GFP_KERNEL without
__GFP_ACCOUNT, allowing unprivileged users to pin kernel memory outside
memcg charging. The shared tcf_exts_init_ex() action array allocation in
cls_api.c was also uncharged; this patch closes it along with the
per-classifier filter-object/percpu/aux allocations that remain
unaccounted.

Add GFP_KERNEL_ACCOUNT to:
- the shared tcf_exts_init_ex() action array (cls_api.c), common to every
filter of every classifier (32 pointers, 256 bytes);
- the filter-object, per-CPU-counter, and per-filter aux allocations in
cls_basic, cls_bpf, cls_cgroup, cls_flow, cls_flower, cls_fw,
cls_matchall, cls_route and cls_u32;
- the u32_init_knode() replace-path knode allocation (cls_u32.c), which
allocates the same struct tc_u_knode + sel.keys on every replace of an
existing knode and was missed by the create-path-only conversion.

Also fix the cls_basic error path: basic_change() inserts fnew into the
IDR before allocating the per-CPU counter. If alloc_percpu() fails the
errout path kfree'd fnew without idr_remove, leaving a dangling pointer
in the IDR. With GFP_KERNEL_ACCOUNT the percpu alloc becomes failable
on demand (memcg at memory.max), making the dead path attacker-reachable
and burning the handle permanently. Add the idr_remove on the percpu
failure path, matching the basic_set_parms failure-path pattern.

Note: vega@nebusec.ai provided a poc for basic_cls, but it was easy to
extend to the other classifiers.

Conditions to recreate the bug:
- CONFIG_NET_SCHED, CONFIG_NET_CLS_* (the classifier being used),
CONFIG_NET_CLS_ACT, CONFIG_MEMCG, CONFIG_USER_NS, CONFIG_NET_NS.
- Unprivileged user in a fresh user+network namespace (unshare -Urn),
or root with CAP_NET_ADMIN.
- Create a large number of tc filters (e.g. tc filter add dev lo
ingress ... <classifier> ...) while watching a memcg-limited cgroup:
system slab grows far faster than memory.current, pinning kernel
memory outside memcg charging.