Detect process exit through the chardev's ->release instead of the sched_process_exit probe #1610

Process death is currently detected by a sched_process_exit probe filtered on is_agnocast_pid() (agnocast_kmod/agnocast_internal.c:239-248), a hand-built substitute for what a chardev ->release gives for free, and it costs an atomic read plus an RCU hash lookup on every process exit on the machine, agnocast-related or not. Driving cleanup from ->release on /dev/agnocast instead (the fops has only .unlocked_ioctl today, agnocast_kmod/agnocast_init.c:19-22) would delete the probe and the filter, and would remove the pid-reuse reasoning that every pid-keyed table in the module currently carries; the ring buffer and worker kthread stay, with ->release replacing the probe as what feeds them.

The blocker is that the file's lifetime is not the process's: agnocast_fd is opened without O_CLOEXEC (src/agnocastlib/src/agnocast.cpp:533) and every daemon child reuses the inherited descriptor rather than reopening (:180, :208, :297), so one struct file is shared by the application and all three of its daemons — a naive .release would never fire for an application process while its unlink daemon, which lives as long as the IPC namespace, still holds it. Making the file per-process (O_CLOEXEC plus a reopen in each child) is therefore a prerequisite.