Fix routing replay split sizes for attention

[vivekkalyan]

Summary

• preserve the full attention token layout when recording routing replay token UID sets
• keep the compacted GDN token UID layout only for GDN replay
• add a unit test that covers the differing attention-vs-GDN replay shapes

Why

The Megatron routing replay path could fail with split_with_sizes expects split_sizes to sum exactly ... when attention replay used compacted GDN token UID sets. Attention needs the original flattened token layout, while GDN uses the compact routed-token layout.

Validation

• uv run --with torch --with safetensors --with megatron-core==0.17.0 --with transformers==5.2.0 --group dev pytest tests/unit/test_moe_routing_replay.py tests/unit/test_dedicated_config.py
• Sky 2x H200 Bonnie Megatron repro against this fix completed 1 training step without the split-size crash
• Stacked LoRA PR smoke run against this branch also completed 1 Megatron training step successfully

[FurtherAI]

Looks good, seems like a misconception from Codex and not caught by tests because they fill the packed seq and don't have padding.