i#2502 locks: Add acquires to bare loads in lock routines

[derekbruening]

Adds load-acquire semantics to a number of bare loads (i.e., not
inside a mutex) in DR's various lock routines.

This includes a key missing load-acquire in the spinlock loop in
d_r_mutex_lock_app() which might help explain some performance issues
such as #4279.

Since some of these are in the utils.h header, and we have ordering
dependencies between utils.h and arch_exports.h, this required
splitting the atomic defines out of arch/arch_exports.h into a new
file arch/atomic_exports.h. Since those atomic defines use ASSERT
which references dynamo_options, this further required splitting the
option struct out of options.h into options_struct.h.

Issue: #2502, #4928, #4279

[derekbruening]

The header code that was moved was not modified. All the not-just-moved modifications (the inserted acquires) are in utils.c and utils.h.

Unfortunately there isn't a simple way to test these things. I ran dr$sim -offline thousands of times on the Jenkins machine and there are no hangs so this doesn't make anything worse; there is still a hang in dr$sim online (#4928). I don't have a setup to easily test performance losses or gains; we'll try to ask on #4279 to do that once this is in to see if it helped there.