fix data race in ReentrantLock fallback for targets without 64bit atomics
See [Zulip](https://rust-lang.zulipchat.com/#narrow/channel/269128-miri/topic/reentrant.20lock.20failure.20on.20musl) for details: the address used to identify a thread might get lazily allocated inside `tls_addr()`, so if we call that *after* doing the `tls_addr.load()` it is too late to establish synchronization with prior threads that used the same address -- the `load()` thus races with the `store()` by that prior thread, and might hence see outdated values, and then the entire logic breaks down.
r? `@joboet`
6e784f842a