feat: bound Megatron checkpoint retention

[skirdey]

Summary

• add opt-in --save-retain-count for long-running RL jobs
• prune only checkpoint directories at or below Megatron latest completed iteration marker
• never remove in-flight directories newer than the marker, symlinks, release checkpoints, or unrelated paths
• support synchronous and asynchronous checkpoint saves

Motivation

Long jobs currently retain every multi-gigabyte Megatron checkpoint unless an external supervisor deletes them. This can exhaust local disks even when the job only needs a small resumable history. The retention operation runs only after synchronous completion, or after maybe_finalize_async_save confirms an asynchronous save.

The option is disabled by default.

Test plan

• PYTHONPATH=$PWD uv run --no-project --with pytest --with pyyaml --with omegaconf pytest -q tests/utils/test_checkpoint_retention.py tests/test_megatron_argument_validation.py
• 24 passed
• python3 -m py_compile slime/utils/checkpoint_retention.py slime/utils/arguments.py slime/backends/megatron_utils/actor.py tests/utils/test_checkpoint_retention.py
• git diff --check