aarch64: IRQ exit blocks on PROCESS_MANAGER under an unmasked try_manager() holder

[ryanbreen]

aarch64: IRQ exit blocks on PROCESS_MANAGER under an unmasked try_manager() holder

Lock: PROCESS_MANAGER (kernel/src/process/mod.rs:146)

Arch: aarch64

Chain:
kernel/src/arch_impl/aarch64/boot.S:699 bl check_need_resched_and_switch_arm64 ->
kernel/src/arch_impl/aarch64/context_switch.rs:5187 (first statement) ->
kernel/src/arch_impl/aarch64/context_switch.rs:5191 crate::task::process_task::drain_deferred_fault_sigsegv_exits() ->
kernel/src/task/process_task.rs:989 ProcessScheduler::handle_thread_exit(tid, -11) ->
kernel/src/task/process_task.rs:764 ->
kernel/src/process/mod.rs:772 blocking crate::process::manager()

Thread-side holders:
The 9 unmasked aarch64 try_manager() holders:
kernel/src/syscall/handlers.rs:3875 sys_fcntl
kernel/src/syscall/signal.rs:1839, :1859, :1902 sys_pause_with_frame_aarch64
kernel/src/syscall/signal.rs:2064 sys_sigreturn_with_frame_aarch64
kernel/src/syscall/signal.rs:2152, :2203, :2247 sys_sigsuspend_with_frame_aarch64
kernel/src/arch_impl/aarch64/syscall_entry.rs:218 check_and_deliver_signals_aarch64

Note: This is distinct from #812 because it sits above the PREEMPT_ACTIVE early return (kernel/src/arch_impl/aarch64/context_switch.rs:5201-5220) and above each from_el0 test, so neither the preempt state nor the interrupted EL gates it. A fix that gates aarch64 do_softirq on the whole preempt_count leaves this chain untouched, because this is not a softirq.

Not measured here. This is a code-path reachability argument with no reproduction.

Related: #812, IRQ-CONTEXT-LOCK-CENSUS-2026-09-05.md section 5.F2