LART: Optimize for TSO if weakest used memory order is acq_rel.

In this case we can flush only first entry in store buffer, and we can bypass
buffer for thread-private stores.