i#2502 locks: Fix bug in ATOMIC_MAX on AArchXX

[derekbruening]

Adds a missing load-acquire in a loop in ATOMIC_MAX that causes hangs
in RSTATS_ADD_PEAK.

Tested on the client.annotation-concurrency test on AArch64 where
without this fix it hangs under release-build plain DR every ~100
runs. With the fix, there is no hang in 20K runs.

Issue: #2502, #4928

[derekbruening]

I am highly confident that this fixes #5366. Without this fix I saw hangs in dr$sim -offline every ~100-300 runs; with the fix, no hangs in 20K runs. That is similar to the plain DR rates. dr$sim online however also has a different, more frequent hang that #4928 covers: have not figured that one out yet.