Fix TslibDataModule setup re-entrancy guards

[Ajeem-git]

Summary

• reuse the original randomized split indices when setup() is called repeatedly
• check the actual dataset attributes instead of nonexistent private names before rebuilding datasets
• add a regression test covering repeated setup(stage="fit") calls

Testing

• python3.11 -m py_compile pytorch_forecasting/data/_tslib_data_module.py pytorch_forecasting/data/tests/test_tslib_data_module.py

Closes #2218

[codecov]

Codecov Report

❌ Patch coverage is 45.23810% with 23 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@af9b14f). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...h_forecasting/data/tests/test_tslib_data_module.py	0.00%	20 Missing ⚠️
pytorch_forecasting/data/_tslib_data_module.py	86.36%	3 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2245   +/-   ##
=======================================
  Coverage        ?   86.53%           
=======================================
  Files           ?      166           
  Lines           ?     9784           
  Branches        ?        0           
=======================================
  Hits            ?     8467           
  Misses          ?     1317           
  Partials        ?        0           

Flag	Coverage Δ
cpu	86.53% <45.23%> (?)
pytest	86.53% <45.23%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Meet-Ramjiyani-10]

Nice work !
Verified the attribute names are now correct (train_dataset vs _train_dataset).

One question: Should stage="test" and stage="predict" also have re-entrancy guards? Currently only "fit" is protected.

Otherwise LGTM.

[Ajeem-git]

Good catch. test was already guarded with a check that only creates the dataset once, but predict was not consistent with that yet. I pushed a small follow-up to make predict re-entrant as well, and added a regression test for repeated setup(stage="predict") calls.

[Meet-Ramjiyani-10]

Cool! The predict stage guard and regression test look correct. LGTM

One small note for future consistency: the test stage still uses the old _test_dataset guard (if not hasattr(self, "_test_dataset"):). Since _test_dataset is never assigned, this guard doesn’t actually prevent re-creation. Might be worth a follow-up PR to align it with the fit/predict pattern. Not a blocker for this PR though.