Use-after-free race in kernel `k_queue_peek_head/tail` due to missing spinlock

Summary

The kernel queue helper z_queue_node_peek() in kernel/queue.c dereferences a node taken from a queue's data_q list, reading the node's flag byte and, for items enqueued via k_queue_alloc_append/alloc_prepend, the data pointer of an internally allocated alloc_node struct. The implementations of z_impl_k_queue_peek_head() and z_impl_k_queue_peek_tail() performed this read-and-dereference without holding the queue's spinlock, while every other accessor of the same list — including k_queue_get(), which unlinks a node and k_free()s its backing alloc_node — operates under that lock.

Because peek was unsynchronized, a concurrent k_queue_get() on the same queue (on an SMP build, or under preemption/ISR concurrency) can free the node between the moment peek obtains the node pointer and the moment it dereferences it. The peek then reads flag bits and a data pointer out of freed, potentially re-allocated heap memory and returns a stale or dangling pointer to its caller. k_fifo and k_lifo are thin wrappers over k_queue, so this affects buffer queues used throughout the net_buf, Bluetooth, USB, and networking subsystems; the peek operations are also system calls reachable from CONFIG_USERSPACE threads.

The consequences are a use-after-free read that can leak stale heap contents (one pointer word) and, when the returned dangling pointer is subsequently consumed as a live buffer, a dereference that can crash the system or corrupt memory. Exploitation requires winning a small race window with local access (e.g. a userspace process racing k_queue_peek_* against k_queue_get on a shared queue, or two CPUs), so practical impact is bounded and of low severity.

The fix wraps both peek implementations with k_spin_lock/k_spin_unlock on the queue lock, making the read-and-dereference atomic with respect to the concurrent unlink-and-free and bringing peek into line with the rest of the queue's locking discipline.