CVE-2026-10653: Non-atomic `net_buf` reference counts cause double-free / free-list corruption under concurrent unref

The Zephyr netbuf library (lib/netbuf/buf.c) manipulated both of its reference counts -- the per-header buf->ref and the per-data-block refcount at the start of each variable/heap data allocation -- with plain non-atomic C operators (buf->ref++, if (--buf->ref > 0), if (--(refcount))).

The API is documented as self-synchronizing: callers may share one buffer across threads (e.g. via kfifo) and each holder independently calls netbufunref() with no surrounding lock. Under true concurrency (SMP, or single-core preemption between the non-atomic load and store while another context unrefs the same buffer), two holders can both observe the same prior reference value and both conclude they are the last reference.

For heap/variable-data pools (mempooldataunref/heapdataunref, used by zbus message subscribers, the IP stack RX/TX buffers when CONFIGNETBUFFIXEDDATASIZE=n, capture, wireguard, ISO-TP and usbip) this produces a double kheapfree()/kfree() of the same block -- heap-metadata corruption and a use-after-free on the heap-hardening poison pattern.

For the per-header refcount the buffer is returned to the pool free LIFO twice for any pool type (including fixed-data pools used by Bluetooth and networking), corrupting the free list so a later allocation hands the same buffer to two owners.

The fix converts both refcounts to atomicinc/atomicdec (overlaying buf->ref in an atomict-sized union and changing the data-block refcount from uint8t to atomict).

Impact is gated on genuine concurrency and on an application architecture that shares one buffer among multiple independent unref'ers; the trigger is a refcount/timing race rather than packet content, so an external attacker has at most weak indirect influence over the race window. Affects all Zephyr releases through v4.4.0.

This fix is not being backported to v3.7-branch (LTS). The backport was attempted and closed unmerged (#111181): the v3.7 networking tree has diverged from main, and the new atomic word-packing -- together with the assertions it adds -- turns pre-existing v3.7-only reference-counting defects elsewhere in the stack into hard faults, so landing the change faithfully would mean pulling an open-ended set of additional v3.7-only fixes into an LTS branch. v3.7 remains affected. Applications on v3.7 that share one netbuf across threads should serialize their own netbufunref() calls rather than rely on the documented self-synchronizing behaviour. The fix is on main and has been backported to v4.3-branch (#110852) and v4.4-branch (#110853).