csky: Fixup asm/cmpxchg.h with correct ordering barrier

Optimize the performance of cmpxchg by using more fine-grained
acquire/release barriers.

Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Paul E. McKenney <paulmck@kernel.org>