[model, conversion] fix: Handle optional DeepSeek V4 config and PP mappings

[bobboli]

Summary

• accept nested rope_theta values in DeepSeek V4 Hugging Face configs
• skip auto-mapped parameters that are absent from every pipeline-parallel stage during Megatron-to-HF export
• add focused regression coverage for both cases

These defensive compatibility changes are separated from the required DeepSeek V4 hash-routing fix. They are not dependencies of #3 and are retained as a draft for independent review.

[bobboli]

Not required by the current DeepSeek V4 direct-HF rollout chain. The optional rope/config and PP mapping branch is retained for possible follow-up work.