The current ARC fetch/return atomics provide fully ordered semantics
only with 2 full barriers around the operation.

Instead implement them as relaxed variants without any barriers and
rely on generic code to generate the fully-ordered, acquire and release
varaints by adding the appropriate full barriers.

This helps elide some extra barriers in case of acquire/release/relaxed
calls.

bloat-o-meter for hsdk defconfig shows codegen improvements, although
numbers below inflated due to unrelated inlining heuristic changes

| bloat-o-meter vmlinux-643babe34fd7-non-relaxed vmlinux-45aa05cb44d7-relaxed
| add/remove: 2/5 grow/shrink: 42/1222 up/down: 4158/-14312 (-10154)
| Function                                     old     new   delta
| ..
| sys_renameat                                 462     476     +14
| ip_mc_inc_group                              424     436     +12
| do_read_cache_page                          1882    1894     +12
| ..
| refcount_dec_and_mutex_lock                  254     250      -4
| refcount_dec_and_lock_irqsave                258     254      -4
| refcount_dec_and_lock                        254     250      -4
| ..
| tcp_v6_route_req                             246     238      -8
| tcp_v4_destroy_sock                          286     278      -8
| tcp_twsk_unique                              352     344      -8

Link: https://lore.kernel.org/r/20180830144344.GW24142@hirez.programming.kicks-ass.netSuggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@kernel.org>