arm64: switch to the IRQ stack during ISR execution

Avoid executing ISRs using the thread stack as it might not be sized
for that. Plus, we do have IRQ stacks already set up for us.

The non-nested IRQ context is still (and has to be) saved on the thread
stack as the thread could be preempted.

The irq_offload case is never nested and always invoked with the
sched_lock held so it can be simplified a bit.

Signed-off-by: Nicolas Pitre <npitre@baylibre.com>