scheduler: harden SIGKILL SMP and clone semantics

scheduler: harden SIGKILL SMP and clone semantics

Fix the remote SIGKILL path so it no longer tears down a task on behalf of another CPU before that task reaches a safe exit point. SIGKILL is now queued as a fatal pending signal, forcibly unblocked, blocked/sleeping tasks are woken, and reschedule IPIs are sent so the target exits on its own CPU.

Fix CLONE_VM address-space accounting by using the global address-space refcount table to detect the first shared-mm clone and to decide teardown during reap. This closes the leak/counter skew when a non-leader thread creates another CLONE_VM thread.

Align clone semantics with the implementation by rejecting unsupported shared-state flags from the syscall interface, while still inheriting cwd, file references, descriptor flags, signal handlers, and signal mask by copy. Update pthread/fulltest users accordingly.

Also prevent userspace from changing SIGKILL/SIGSTOP disposition or blocking them, make default SIGKILL delivery fatal in usermode signal dispatch, and extend the kill regression test so the child tries to block SIGKILL before spinning.