Megatron offload_after_refit leaves params on GPU (re-onload via offload_before_refit)

[yfw]

MegatronPolicyWorker.offload_after_refit() leaves params on GPU after a colocated refit, instead of offloading it. Root cause is in #2530:

• feat: PPO with MCore #2530 changed offload_before_refit() to unconditionally onload params to GPU (move_model("cuda", move_params=True, ...)) — needed because PPO's new finish_training() offloads params before refit, and streaming needs them back.
• But offload_after_refit() reuses offload_before_refit() as its final step (to drop optimizer/grads/caches). So it re-onloads the params it just moved to CPU.

Impact / how it surfaced

Caught by test_vllm_megatron_weight_update_memory (assert current_allocated <= 0.1 → observed 1884.59). The test is mcore + deselected as expensive, so it's latent on main and only runs on gates that include expensive tests. DTensor is unaffected (no finish_training; its offload_before_refit is pure-offload and lands at 0 GB).

Notes:

The onload is only needed for megatron + PPO (after finish_training()); it's a no-op in GRPO (params already resident). Having offload_before_refit onload params seems a bit counterintuitive. Can we instead explicitly re-onload only in this case (PPO?)