fix: warn when training batches are fewer than accumulate_grad_batches

[giacomo-ciro]

What does this PR do?

When a training epoch contains fewer batches than accumulate_grad_batches, Lightning still forces an optimizer step on the last batch of the epoch (a deliberate design choice from #2853, meant to flush leftover gradients when the epoch ends mid-accumulation-cycle). If the whole epoch is shorter than accumulate_grad_batches, every batch is the "last batch," so this rule fires on every step and gradient accumulation never actually happens as configured, gradients end up accumulated over fewer batches than intended, with no error or log to indicate it.

As agreed with @deependujha in the issue discussion, this PR adds a PossibleUserWarning (instead of a docs-only fix) raised in FitLoop.setup_data() when num_training_batches < accumulate_grad_batches, following the same pattern as the existing log_every_n_steps warning in the same method.

Fixes #21808

This PR does not introduce any breaking changes.

Before submitting

• Was this discussed/agreed via a GitHub issue? Yes, see accumulate_grad_batches is silently ignored when epoch length < accumulate_grad_batches, because Lightning always steps on the last batch #21808.
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? Not needed. Per @deependujha 's preference, we went with a runtime warning instead of a documentation update.
• Did you write any new necessary tests? Yes, added test_warning_with_dataloader_shorter_than_accumulate_grad_batches.
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request? None.
• Did you update the CHANGELOG? Yes.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79%. Comparing base (7b57f9c) to head (7454080).
⚠️ Report is 3 commits behind head on master.
✅ All tests successful. No failed tests found.
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

❗ There is a different number of reports uploaded between BASE (7b57f9c) and HEAD (7454080). Click for more details.

HEAD has 158 uploads less than BASE

Flag	BASE (7b57f9c)	HEAD (7454080)
cpu	80	42
python	5	3
lightning_fabric	26	0
pytest	41	0
python3.12	23	12
python3.13	16	9
python3.12.7	18	9
python3.10	6	3
lightning	30	15
python3.11	12	6

Additional details and impacted files

@@            Coverage Diff            @@
##           master   #21811     +/-   ##
=========================================
- Coverage      87%      79%     -8%     
=========================================
  Files         270      267      -3     
  Lines       24067    24010     -57     
=========================================
- Hits        20899    18865   -2034     
- Misses       3168     5145   +1977     

[giacomo-ciro]

@deependujha are the failing checks expected or do I need to do something about it?

[giacomo-ciro]

@deependujha all checks had passed. I merged master again because it was updated in the meantime.

[giacomo-ciro]

@deependujha shall we merge?

[giacomo-ciro]

@deependujha I know you have a lot of PRs on your plate! When you have a moment, could you please approve the workflows so we can merge this? The fix is minimal and won't affect any other tests.