Add debug log before training loop starts

[mvpatel2000]

What does this PR do?

Streaming often takes a while to start first batch. This changes last seen debug message from user to starting training loop. This avoids confusion with users thinking spinning dataloaders is taking a really long time.

[bigning]

simple change, lgtm.

[dblalock]

Bonus points: can we make batch 0 print for any console_log_interval? So that with an interval of 10 we print after batches {0, 10, 20}, etc instead of {9, 19, 29}, etc? I think this would address 80% of the motivation for the more granular logging I mentioned in the JIRA issue. (May or may not make sense to do in this PR.)

[mvpatel2000]

Bonus points: can we make batch 0 print for any console_log_interval? So that with an interval of 10 we print after batches {0, 10, 20}, etc instead of {9, 19, 29}, etc? I think this would address 80% of the motivation for the more granular logging I mentioned in the JIRA issue. (May or may not make sense to do in this PR.)

I'll follow up offline. We already do this, maybe you were testing a resumption?