Fix trainer.test() logging incorrect epoch from checkpoint (#20052)

[devraj2307]

What does this PR do?

Fixes #20052

Restores the checkpoint epoch and global step when running Trainer.validate(), Trainer.test(), or Trainer.predict() with an explicit ckpt_path.

Previously, these entry points restored the model weights from the requested checkpoint but left the fit-loop progress counters unchanged. As a result, hooks and loggers still
observed the last in-memory training epoch/step instead of the values stored in the selected checkpoint.

This PR:

• restores the fit-loop progress needed to back trainer.current_epoch and trainer.global_step for evaluation/prediction checkpoint restores
• adds a regression test covering trainer.test(ckpt_path=...) with an older checkpoint
• updates restore-path expectations to reflect the corrected global_step behavior during evaluation
• adds a changelog entry

No breaking changes.

---

📚 Documentation preview 📚: https://pytorch-lightning--21714.org.readthedocs.build/en/21714/