Fix model convert when use latest megatron

[alexqdh]

• Add --use-gated-attention compatibility to the HF-to-torch_dist conversion tool while retaining --attention-output-gate.
• Default enable_gloo_process_groups to True when it is unavailable in the Megatron argument namespace.
• Support both old and new Megatron tokenizer utility import paths.
• Allow the custom Qwen3.5 and Qwen3-Next attention modules to accept Megatron’s optional name argument.