[model] fix: Map DeepSeek V4 hash layer config

[bobboli]

Summary

• map the DeepSeek V4 hash-layer count to the Megatron-Core dsv4_n_hash_layers field
• read the same field when exporting the Hugging Face config
• add import and export regression coverage

The previous code assigned moe_n_hash_layers dynamically. The Megatron-LM runtime defines the hash-layer count as TransformerConfig.dsv4_n_hash_layers and reads it in Router._init_routing_mode() to select hash routing. It does not read moe_n_hash_layers.

As a result, the trainer constructed learned routers for the first three layers while SGLang used deterministic hash routing. That caused both the R3 replay failure and the large rollout/trainer mismatch.

Verification

• uv run python -m pytest tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py -q: 16 passed
• targeted pre-commit hooks: passed