fix: check IS_ERR(wg) before using the device in wg_set_device

[ExzoTikStyle]

wg_set_device() initialises has_protection from awg_has_header_protection(wg)
in the declaration block, above the IS_ERR(wg) check. When the request names an
interface that doesn't exist, lookup_interface() returns an ERR_PTR and the
helper dereferences it right away.

The oops lands inside down_read(), and genl_rcv_msg() holds the genl family
lock across the doit callback, so the lock is never released. Everything on the
amneziawg family blocks in uninterruptible sleep from then on — including
awg setconf for tunnels that were up and working — and genl_unregister_family()
hangs too, so the module can neither be used nor unloaded. Only a reboot clears it.
rmmod -f isn't a way out either, since stock Ubuntu kernels build without
CONFIG_MODULE_FORCE_UNLOAD.

BUG: kernel NULL pointer dereference, address: 000000000000050d
Oops: 0002 [#1] PREEMPT SMP NOPTI
RIP: 0010:down_read+0x1e/0xc0
Call Trace:
 <TASK>
 awg_has_header_protection+0x19/0x50 [amneziawg]
 wg_set_device+0x43/0xe50 [amneziawg]
 ? genl_family_rcv_msg_attrs_parse.constprop.0+0x93/0x100
 genl_family_rcv_msg_doit+0xfa/0x160
 genl_family_rcv_msg+0x189/0x260
 ? __pfx_wg_set_device+0x10/0x10 [amneziawg]
 genl_rcv_msg+0x4c/0xb0
 genl_rcv+0x28/0x50
 </TASK>
note: awg[80691] exited with irqs disabled
note: awg[80691] exited with preempt_count 1

CR2 being 0x50d rather than a near-zero address is the ERR_PTR giveaway:
(void *)-19 plus the offset of header_protection.lock in struct wg_device.

The call is the left operand of the ||, so it runs whether or not
WGDEVICE_A_HEADER_PROTECTION_KEY was passed — a plain
awg set doesnotexist listen-port 51820 reaches it too. Introduced in #192.

The first commit moves the check above the use. It also adds an
IS_ERR_OR_NULL() guard inside awg_has_header_protection(), since it takes a
device pointer from several call sites.

The second commit is unrelated to the crash but in the same file:
awg_header_protection_set_key() writes p->key[] and p->has_protection
while holding only down_read(). Concurrent setters can interleave, and
awg_header_protection_init() can read a half-updated key while building the
chacha state — silently wrong header protection rather than a reported error.
Happy to split it out if you'd rather keep this PR to the crash.

Verification

Built and run against 6.8.0-136-generic (Ubuntu 22.04.5), on the same host
where the oops was originally hit.

Before, on 3.0.20260731-04 from ppa:amnezia/ppa:

# awg set doesnotexist header-protection-key /dev/null
Killed
# echo $?
137
# awg-quick up awg0        # hangs, unkillable
# ps -eo pid,stat,wchan:24,cmd | awk '$2 ~ /D/'
  81680 D    genl_rcv_msg             awg setconf awg0 /dev/fd/63
  82328 D    genl_unregister_fami     rmmod amneziawg

After, with this branch:

# awg set doesnotexist header-protection-key /dev/null
Unable to modify interface: No such device
# echo $?
1
# awg set doesnotexist listen-port 51820
Unable to modify interface: No such device
# echo $?
1
# ps -eo pid,stat,cmd | awk '$2 ~ /D/' | wc -l
0

dmesg is clean, and an existing AWG 3 tunnel (with HeaderProtectionKey set)
comes back up on the patched module and passes traffic as before.