Training crash with TE CUDA graphs

# Training crash with TE CUDA graphs (`--cuda-graph-scope attn`) after ~89 iterations on Qwen3-235B

## Environment

| Component | Version |
|---|---|
| Megatron-Core | `0.19.0+ddaa315fe` |
| PyTorch | `2.8.0a0+5228986c39.nv25.6` |
| Transformer Engine | `2.16.0` |

## Configuration

- Qwen3-235B-A22B, TP=2, PP=8, VPP=4, EP=8, ETP=1, DP=16
- BF16
- MoE alltoall dispatcher
- `--overlap-moe-expert-parallel-comm`
- CUDA graph: `--cuda-graph-impl transformer_engine --cuda-graph-scope attn`

## What Happens

Training runs normally for ~80 iterations, then grad norm starts oscillating wildly between normal values and tens of millions, before hitting NaN at iteration 91 and crashing:

| Iteration | Grad Norm |
|---|---|
| 80 | 91 |
| 81 | 463,582 |
| 82 | 50 |
| 83 | 21,210,958 |
| 87 | 127 |
| 88 | 27,622,466 |
| 89 | 11,356,515 |
| 91 | NaN → crash |

The crash message is:

```
ERROR:megatron.core.rerun_state_machine: found NaN in local grad norm for bucket #0
in backward pass before data-parallel communication collective
```

## Workaround

Removing `--cuda-graph-impl transformer_engine --cuda-graph-scope attn` makes the crash disappear.

## Question

Has anyone seen this before, or knows what might be causing it?