Support grouped expert bias in checkpoint split/merge adapters

Summary

Add an explicit checkpoint conversion contract for grouped expert MLP bias tensors.

The shared MoE adapter currently converts grouped expert weights, but not gate_up_proj_bias / down_proj_bias. Per-expert HF bias keys can therefore pass through without being regrouped into the native layout. #3610 makes this safe by disabling its low-memory DCP capability for expert_bias=True and rejecting conversion clearly instead of silently loading or exporting incorrect keys.

Scope

• Define the HF per-expert bias keys and native grouped bias layouts for gated and non-gated experts.
• Implement both HF-to-native merge and native-to-HF split conversion.
• Preserve rank-local expert ownership under EP/DTensor configurations.
• Cover ordinary grouped experts and the maintained TE/DeepEP-backed configurations that expose expert bias.
• Keep router correction-bias tensors separate; they already have model-specific rename handling and are not this issue.

Acceptance criteria

• Load, save, and load-save-reload tests cover expert weights and both expert bias tensors.
• Missing or partial bias keys fail loudly.
• Single-device and representative EP layouts produce the same values as a trusted HF-format reference.
• The low-memory DCP capability is enabled only for runtime layouts whose destinations are model-backed or otherwise memory-bounded.

Parent: #3576

Follow-up to #3610.