fix(train): fail fast on incompatible DLRover flash checkpoint APIs

[Excelius-Wang]

Summary

• validate the DLRover Flash Checkpoint API before loading the model
• fail fast with an actionable source-install command when blocking or max_steps is unavailable
• add regression tests for the released DLRover 0.6.1 API and the current source API

Motivation

The latest published DLRover release, 0.6.1, does not accept the blocking and max_steps arguments used by ms-swift. The incompatibility is currently discovered only while saving a checkpoint or shutting down training.

Silently dropping these arguments would be unsafe because blocking=True prevents the final checkpoint from being skipped while a previous asynchronous save holds the shared-memory lock. This change keeps the current checkpoint semantics and reports the incompatible dependency before model loading.

Validation

• focused unit tests: 2 passed
• project test runner: 2/2 passed
• all relevant pre-commit hooks passed
• git diff --check passed

Related to #9912.