fix: initialize dense local CUDA graph state

[nvegesna-netizen]

Problem

GraphableMegatronModule can create a local CUDA graph manager from its constructor. For an MLP-only graph scope, that path reads is_moe_layer before TransformerLayer initializes it, so dense layers fail during construction.

Fix

Initialize the dense default before the base constructor while preserving the value already set by MoETransformerLayer. TransformerLayer still derives the final value from the constructed MLP later in initialization.

Validation

• Added a focused unit regression that constructs a dense TransformerBlock with local MLP CUDA graph scope and verifies every layer receives a graph manager.

[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.