Track Qwen3.5-related issues

[zpqiu]

• MCore Path
  • Add / track CP support: [main] feat(moe): Support packed sequence for gated delta net (GDN) NVIDIA/Megatron-LM#2645
• AutoModel Path
  • Move the FLA dependency from the dev group ([dependency-groups]) to optional extras ([project.optional-dependencies]) so that NeMo-RL can install it downstream via pkg[extra]. If FLA is not installed
    • No CP support
    • Worse performance
    • build: move flash-linear-attention back to optional-dependencies Automodel#1894
  • Fix the default config path where Torch Adam is used without FP32 master weights, as this can slow down convergence.
    • TE FusedAdam can be used as a workaround.
    • AutoModel should correctly support / apply the FP32 master weight setting. fix: fp32 master weights for custom MoE models under FSDP2 Automodel#1896