refactor(checkpoint-object-manager): deprecate C++ object_manager layer for lightweight Python subprocess async deletion

[Leahlijuan]

Refactor: deprecate C++ object_manager layer for lightweight Python subprocess async deletion

This patch modernizes the ML-Flashpoint directory deletion lifecycle by fully replacing the native C++ object_manager module with streamlined subprocess.Popen calls. This drastically simplifies the codebase overhead during checkpoint cleanup routines.

[g-husam]

hm ideally we'd at least return something that informs the caller there was an error. Otherwise they can't distinguish between "nothing to do" and "failed to delete".

Maybe bubble up the exception? Or can return a wrapper type encapsulating an Optional proc and an optional error e.g.

@dataclass
class DeletionResult:
    proc: Optional[subprocess.Popen] = None
    error: Optional[Exception] = None
    skipped: bool = False

Not a blocker, but will make the API more explicit.

Then in the failure test, we can confirm there was an error by asserting on the result, as opposed to just expecting proc to be None, which is the same expectation when there is no error and nothing to do.

[github-actions]

Python Code Coverage Summary

Package	Line Rate	Branch Rate	Health
src.ml_flashpoint	100%	100%	✔
src.ml_flashpoint.adapter	100%	100%	✔
src.ml_flashpoint.adapter.megatron	97%	95%	✔
src.ml_flashpoint.adapter.nemo	98%	93%	✔
src.ml_flashpoint.adapter.pytorch	99%	92%	✔
src.ml_flashpoint.checkpoint_object_manager	93%	93%	➖
src.ml_flashpoint.core	95%	92%	✔
src.ml_flashpoint.replication	81%	81%	❌
Summary	95% (2390 / 2521)	92% (571 / 624)	➖

Minimum allowed line rate is 90%

[github-actions]

C++ Code Coverage Summary

Package	Line Rate	Branch Rate	Health
src.ml_flashpoint.checkpoint_object_manager.buffer_object	93%	54%	✔
src.ml_flashpoint.replication.transfer_service	79%	41%	❌
Summary	82% (899 / 1097)	43% (668 / 1539)	➖

Minimum allowed line rate is 80%