More places might need memory barriers on AArch64

[HertzDevil]

#13010 covers just one place where memory barriers are needed on AArch64. The following spec fails randomly on an M2 when -Dpreview_mt is provided:

crystal/spec/std/mutex_spec.cr

Lines 11 to 28 in 2d49b07

	it "works with multiple threads" do
	x = 0
	mutex = Mutex.new
	fibers = 10.times.map do
	spawn do
	100.times do
	mutex.synchronize { x += 1 }
	end
	end
	end.to_a
	fibers.each do |f|
	wait_until_finished f
	end
	x.should eq(1000)
	end

This is probably due to Mutex using not just one, but two Atomics directly. In particular, #unlock calls @state.lazy_set, which is non-atomic.

Some other places that look suspicious:

• The recently added Crystal::AtomicSemaphore, which is part of Add Process.on_interrupt #13034; as of now the type is used only on x86-64 Windows
• Crystal::RWLock, which too calls Atomic#lazy_set