fix(checkpoint): honor train_time_interval under manual_optimization

[wietzesuijker]

What does this PR do?

ModelCheckpoint(train_time_interval=...) silently no-ops under automatic_optimization=False. The manual-opt branch in on_train_batch_end only inspects every_n_train_steps, so a callback configured with only a time interval never fires mid-run. last.ckpt still appears at end-of-fit via on_train_end, which masks the bug at the file level but breaks any workflow that depends on mid-run saves (chained SLURM segments, spot/preempt training).

Surfaced debugging a manual-opt GAN run on chained 3h SLURM segments where every segment threw away the previous one's work. No prior issue filed.

Fix

Mirror the auto-opt branch's skip_batch + skip_time logic inside the manual-opt branch, reusing the existing _train_time_interval / _last_time_checked / trainer.strategy.broadcast plumbing. Pre-optimization-state warning path unchanged.

Test

test_model_checkpoint_manual_opt_train_time_interval installs a spy callback that observes _last_global_step_saved after each batch. End-of-fit file checks miss the bug because on_train_end saves once anyway. Verified test fails on master, passes with the patch. Broader checkpoint suite (test_model_checkpoint.py, test_checkpoint_callback_frequency.py, test_model_checkpoint_manual_opt.py) all green locally.

Before submitting

• Read contributor guideline
• PR does one thing
• Tests pass locally
• CHANGELOG entry added under unreleased Fixed. No breaking changes.

---

AI (Claude) supported my development of this PR.

[justusschock]

we should probably always do that. independent of if train_time_interval is None. worst case we broadcast false but at least that way we prevent a case where some ranks could theoretically hang because they don't expect the broadcast operation (i.e. if code gets out of sync or so)