Conversion of Nemotron 3 Super left out a lot of weights

**Describe the bug**

After conducting GRPO on Nemotron 3 Super, the conversion of weights format from Megatron format to HuggingFace format leaves out lots of weights not converted.

From this screenshot , you can see some weights are not converted:

Full log: https://gist.github.com/soonchangAI/0d60c46178b05b4ce66043e6a2f05185

**Steps/Code to reproduce bug**

Container name: `nvcr.io/nvidia/nemo-rl:v0.5.0.nemotron_3_super`
Conversion script:
```
CKPT=/path/to/checkpoint/step_X0
uv run --extra mcore torchrun --nproc_per_node=8 --nnodes=1 examples/converters/convert_megatron_to_hf_super.py \
--config $CKPT/config.yaml \
--megatron-ckpt-path $CKPT/policy/weights/iter_0000000 \
--hf-ckpt-path $CKPT/hf \
--tp 8 \
--pp 1 \
--ep 8 \
--etp 1
```

**Expected behavior**

Successful conversion of weight format without leaving out any weights