feat(megatron): expose training CUDA Graph config

[seonjinn]

What

• Expose cuda_graph_modules and cuda_graph_warmup_steps in the NeMo-RL Megatron config.
• Forward them to Megatron-Core before TransformerConfig.__post_init__().
• Add disabled-by-default example configs and focused tests.

Why

NeMo-RL already forwards cuda_graph_impl, but ignored these two existing Megatron-Core options when they were set in YAML.

Configuration

policy:
  megatron_cfg:
    cuda_graph_impl: transformer_engine
    cuda_graph_modules: [attn, mlp]
    cuda_graph_warmup_steps: 3

With cuda_graph_impl: "none", module scopes and warmup are inactive; 3 matches the Megatron-Core default.

[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.

[seonjinn]

/claude review