[feature] Support SequentialMLP (non-fused MoE) AutoBridge Mappings for P0 models

User problem

By default when a new model is supported in AutoBridge, only GroupedLinear (moe_grouped_gemm=True) HF <-> Megatron mappings are added.

In ModelOpt Quantization / Pruning, we often need to use model initialized with non-fused MoE (moe_grouped_gemm=False) and need to submit a PR for enabling the mappings in per-model AutoBridge.

Requesting to have P0 important models already supported

• NemotronH
• Qwen3/3.5/3.6
• GPT-OSS
• DeepSeek
• GLM

Some PRs already merged in the past regarding the same:

• https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/2358
• https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/3628
• https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/4499
• https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/4706
• https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5140

Desired outcome

Models can be imported from HF to Megatron with moe_grouped_gemm=False in the transformer_layer_spec / provider.

Alternatives considered

No response

Affected area

area:model

Urgency / use case

Important but not blocking

Extra context

No response

Start by reviewing the related PRs #2358, #3628, #4499, #4706, and #5140 to understand the existing per-model AutoBridge mappings. Cover NemotronH, Qwen3/3.5/3.6, GPT-OSS, DeepSeek, and GLM so HF imports work with moe_grouped_gemm=False in the transformer_layer_spec or provider.