optimizer_cuda_graph=True crashes during graph capture when grad_norm_skip_threshold uses default float('inf')

**Environment:**
1.Megatron-LM: core_r0.18.0 && Transformer Engine: 2.14

**Summary:**
When --optimizer-cuda-graph is enabled, training crashes during the CUDA graph capture step (warmup step) with:
torch.AcceleratorError: CUDA error: operation not permitted when stream is capturing
The error occurs at the grad_norm > grad_norm_skip_threshold comparison in ChainedOptimizer.step().

**Root Cause**
Two commits interact badly:
3d87bfc1b — introduced OptimizerCudaGraphWrapper, which wraps the entire ChainedOptimizer.step() inside a torch.cuda.graph() capture context on a non-default stream.
180131620 — added the grad_norm > grad_norm_skip_threshold check inside ChainedOptimizer.step().

During CUDA graph capture, grad_norm returned by get_grad_norm_fp32() is a 0-dim CUDA tensor (not a Python float) when Transformer Engine is installed, because multi_tensor_scale_tensor_impl is not None and total_norm.pow(...) returns a tensor:

[clip_grads.py#L139-L142](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/optimizer/clip_grads.py#L139-L142)

if multi_tensor_scale_tensor_impl is not None:
total_norm = total_norm.pow(1.0 / norm_type) # ← 0-dim CUDA tensor
else:
total_norm = total_norm.item() ** (1.0 / norm_type) # ← Python float
Then at the comparison:

[optimizer.py#L1482](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/optimizer/optimizer.py#L1482)

if grad_norm > optimizer.config.grad_norm_skip_threshold:
tensor > float('inf') produces a BoolTensor, and Python's if calls .item() → cudaDeviceSynchronize() — which is not permitted while the stream is capturing.

This happens even with the default value float('inf') where the comparison is always False.