fix(megatron): restore saved scheduler progress without accumulation

[hershg]

What does this PR do?

Restore the saved Megatron scheduler counter when loading a checkpoint into an already-used worker. Core loads progress by adding the saved step count; reset the existing counter first so repeated restores do not accumulate steps.

The runtime change is two lines. A regression exercises the production loader with the real Core scheduler, advances it past the saved state, and restores twice. The equipped Megatron CI lane runs that test.

Validation

uv run --isolated --no-project ../cpu-env/bin/python -m pytest -q tests/backends/skyrl_train/distributed/test_scheduler_restore.py

1 passed. Ruff, Black, secret checks, and independent source review passed. The production fix and test match the deployed GLM image source 6538c46a; this PR makes no GLM GPU qualification claim.