Patches for remote sync waiting and log_local checkpointing

[jeffreywpli]

Three main changes (originally motivated by making it easier to chain together consecutive calls to open_lm.main)

1. Allow for user to specify their own custom NCCL timeout. The default is 10 minutes which may not be enough to run final remote sync in all scenarios (causing a timeout because all other processes hit a barrier).

2. Remove unnecessary waits during the initial "test call" of remote sync and final sync by hard-passing in 0 instead of args.sync_every into remote_sync_with_expon_backoff. These are the only instances in main where remote_sync_with_expon_backoff is called directly (as opposed to in a subprocess) and therefore causes whole script to hang when sync_every is not 0.

3. args.log_local perhaps does not work as originally intended, since many logging calls are still gated behind an if is_master(args) instead of if is_master(args, local=args.log_local).