fix(dev): correct params->parameters typo in ChainedOptimizer.step()

[Phlip79]

Summary

Port of #4277 to the dev branch.

Dev hits the same NameError that #4277 fixed on main:

File "megatron/core/optimizer/optimizer.py", line 1380, in step
    and getattr(params[0], "__fsdp_param__", False)
NameError: name 'params' is not defined

The surrounding clip_grad_by_total_norm_fp32 call already binds the local parameters, so this is a pure typo fix.

Changes

• megatron/core/optimizer/optimizer.py — params[0] → parameters[0] in ChainedOptimizer.step() (the Megatron-FSDP + use_precision_aware_optimizer branch).
• tests/unit_tests/distributed/megatron_fsdp/test_mcore_fully_sharded_data_parallel.py — add fp8="hybrid" to the optim_grads_params_fused_adam_e2e param block so the FSDP e2e test actually exercises the fixed code path.

The @pytest.mark.flaky_in_dev decorator is intentionally left in place on dev — whether the typo was the sole cause of flakiness is unclear, so the safeguard is preserved here even though #4277 removed it on main.

Test plan

• CI green on dev
• test_compatible_with_nd_parallel[optim_grads_params_fused_adam_e2e-*] no longer raises NameError

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.