Consider newest heartbeat to compare timeout (#1728)
Fixes https://github.com/pytorch/xla/issues/1690 taking care of case where:
```
t_0 = 0 last mark_step in epoch 1
t_1=t_0+1 start checkpointing (after epoch 1)
t_2=t_1 rendezvous call blocking on others
t_3=t_2+uneven_heartbeat_timeout checkpointing done + rendezvous returns
t_4=t_3+1 mark_step only by worker 0 (w_0)
t_5=t_4+1 check heartbeat process starts -> errors with unhealthy
since for all w_i, s.t. i!=0, w_i.last_heartbeat = t_0 and
t_4 - t_0 = uneven_heartbeat_timeout + 2 > uneven_heartbeat_timeout
So at t_3 we reset w_i.last_heartbeat = now for all i.
```
