fix: overfit_batches uses same batch for train and val

[ved1beta]

Here's a commit message in the requested format for our changes:

What does this PR do?

This PR fixes the issue where overfit_batches=1 uses different batches for training and validation. It ensures that the same batch is used for both training and validation steps when overfitting.

Fixes #15021

Key Changes:

• Modified _resolve_overfit_batches to use the same batch for both training and validation
• Added comprehensive test to verify identical batches are used
• Added TensorFlow and TPU-related environment variables to test allowlist
• Fixed test failures related to environment variable leaks

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
  • Yes, discussed in Overfit batches parameter gives a validation batch #15021
• 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? (if necessary)
  • No documentation changes needed as this fixes a bug in existing functionality
• Did you write any new necessary tests? (not for typos and docs)
  • Yes, added test to verify identical batches are used
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
  • No breaking changes
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)
  • Yes, will add entry for overfit_batches fix

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

---

📚 Documentation preview 📚: https://pytorch-lightning--20731.org.readthedocs.build/en/20731/

[adosar]

Intuitively, we expect overfitting to occur when the model performs well on training set but not on validation. Using the same data for both training and validation essentially tell us only what happens on the training data. However, this is only half the story to properly assess whether overfitting occurs.

@ved1beta Could you please elaborate more on the motivation behind using the same batches for both training and validation?

[ved1beta]

Using the same batch for both training and validation with overfit_batches=1:

-Tests a model's memorization capacity as a debugging technique
-Verifies model implementation correctness by ensuring it can achieve near-zero loss
-Serves as a sanity check before investing time in full dataset training