[wpyaoxueros]

What does this PR do?

When resuming with accumulate_grad_batches > 1, ModelCheckpoint starts with an empty last-saved-step marker even though the restored trainer is already at a checkpointed global_step. The first resumed batch can therefore save that same step again before an optimizer update, overwriting the checkpoint with later loop progress. Initialize the guard from the restored trainer step when a checkpoint was loaded.

Fixes #22033

Validation

• The regression test fails on the base revision: the resumed checkpoint changes from (global_step=2, batches_completed=3) to (2, 4).
• After the fix, the checkpoint remains (2, 3) and the original CPU-only reproduction shows weights resumed from the saved checkpoint matching uninterrupted training.
• pytest --confcutdir=tests/tests_pytorch/callbacks -q tests/tests_pytorch/callbacks/test_model_checkpoint_edge_cases.py: 3 passed.
• Ruff 0.17.1 checks and formatting passed for both changed code files.
• The PyTorch changelog entry is included in this PR.
• The full repository suite was not run.

AI disclosure: Codex assisted with reproducing the bug, implementing the fix, and writing the regression test.