fix: preserve incomplete flash checkpoints

[LeterTao]

Problem

When Flash Checkpoint is enabled, Swift writes checkpoint metadata such as
trainer_state.json before calling DLRover to persist the checkpoint.

If DLRover returns success=False, it may only indicate that asynchronous
persistence has not completed yet. However, Swift currently removes the entire
checkpoint directory in this case, deleting trainer_state.json and other
already-written files.

As a result, checkpoint recovery may fail even though the checkpoint is still
being persisted asynchronously.

Fix

Stop recursively deleting the checkpoint directory when DLRover reports that
the latest checkpoint is not finished.

The existing dlrover_latest.txt mechanism remains responsible for identifying
the latest complete checkpoint. Incomplete checkpoint directories are preserved
but are not selected during resume.

Validation

• python -m compileall -q swift/trainers/mixin.py
• git diff --check

[tastelikefeet]

Hi, thanks for your contribution! Can you check if this PR works in your scenario:https://github.com/modelscope/ms-swift/pull/9990/changes#diff-52addf56cb07f73a14e0e9c60c0f533c601ff29e09b7e22b8540fc93bb851d70R727