[BUG] Fix stale attention metadata after Eagle pre-planner pads the batch

[1096125073]

Problem

When Eagle speculative decoding + DP-attention (global_num_tokens_cpu is set) are both enabled, flash_mla_with_kvcache raises "num_splits must have shape(b+1)" with the following sequence:

1. Eagle pre-planner calls init_forward_metadata with the unpadded batch
2. mark_forward_metadata_ready(False) prevents re-init
3. prepare_mlp_sync_batch pads batch_size in-place — but the stored forward_metadata still has num_splits shaped for the old unpadded batch
4. The attention kernel receives q with the padded batch but num_splits with the unpadded shape → crash

Fix

In _forward_raw (model_runner.py), after prepare_mlp_sync_batch changes the batch, detect the staleness and clear forward_metadata_ready to let init_forward_metadata re-run with the final padded batch:

if (
    forward_batch.forward_metadata_ready
    and not forward_batch.forward_metadata_replan_equivalent
    and forward_batch.batch_size != forward_batch.forward_metadata_planned_bs
):
    forward_batch.forward_metadata_ready = False











<!-- pr-states:start -->
---
### CI States

Latest PR Test (Base): <!-- slot:pr-test:start -->:x: [Run #28145748899](https://github.com/sgl-project/sglang/actions/runs/28145748899)<!-- slot:pr-test:end -->
Latest PR Test (Extra): <!-- slot:pr-test-extra:start -->:x: [Run #28145748780](https://github.com/sgl-project/sglang/actions/runs/28145748780)<!-- slot:pr-test-extra:end -->
<!-- pr-states:end -->

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[1096125073]

@merrymercy Hi, could you help to review this fix? Thanks

[1096125073]

/tag-and-rerun-ci