[bugfix] Support MIXED forward mode in TBO splitter for DP attention

[ch-wan]

Motivation

`--enable-dp-attention` together with `--enable-mixed-chunk` crashes
deterministically at the first chunked-prefill iteration that merges a
running decode. The combination is silently broken — there's no validation
in `server_args.py` rejecting it, but no scheduler iteration with
`forward_mode = MIXED` ever survives long enough to reach the model.

Under DP attention, `prepare_mlp_sync_batch_raw` (`scheduler_dp_attn_mixin.py:187`)
unconditionally calls `TboDPAttentionPreparer.prepare_all_gather`, which
calls `compute_split_seq_index` on the local batch's `forward_mode` —
even when `--enable-two-batch-overlap` is off. That function (and
its companion `compute_split_token_index`) only handled
`EXTEND / TARGET_VERIFY / DECODE / IDLE / PREBUILT` and raised
`NotImplementedError` for `ForwardMode.MIXED`.

Repro on the failing combination:

```
File "python/sglang/srt/managers/scheduler_dp_attn_mixin.py", line 187, in prepare_mlp_sync_batch_raw
local_can_run_tbo, local_forward_mode = tbo_preparer.prepare_all_gather(local_batch)
File "python/sglang/srt/batch_overlap/two_batch_overlap.py", line 398, in prepare_all_gather
self.local_tbo_split_seq_index = compute_split_seq_index(
File "python/sglang/srt/batch_overlap/two_batch_overlap.py", line 94, in compute_split_seq_index
raise NotImplementedError()
NotImplementedError
```

Modifications

Treat `ForwardMode.MIXED` like `ForwardMode.EXTEND` in both
`compute_split_seq_index` and `compute_split_token_index`. After
`mix_with_running`, the running-decode reqs are appended to
`extend_lens` as length-1 entries, so `_split_extend_seqs` and the
cumulative-sum split logic apply unchanged.

Two-line fix in `python/sglang/srt/batch_overlap/two_batch_overlap.py`.
Added a regression test `TestDPAttentionMixedChunk` in
`test/registered/distributed/test_dp_attention.py` that exercises
`--enable-dp-attention --enable-mixed-chunk --chunked-prefill-size 256`
on the existing 2-GPU CI suite.

Out of scope: `--enable-two-batch-overlap` + `--enable-mixed-chunk`
is still unsupported. `OperationsStrategy.init_new_tbo` →
`compute_moe*_layer_operations_strategy_tbo` in
`batch_overlap/operations_strategy.py` (lines 84, 163, 240) also raise
`NotImplementedError` on MIXED. Non-TBO setups never reach those paths
because `can_run_tbo` requires `enable_two_batch_overlap=True`. Happy
to extend in a follow-up if there's interest.

Accuracy Tests

DeepSeek-Coder-V2-Lite-Instruct, `--tp 2 --enable-dp-attention --dp 2 --enable-mixed-chunk --chunked-prefill-size 256`,
gsm8k 200 examples / 128 threads:

```
gsm8k_score = 0.795 (threshold: 0.6)
output_throughput = 1138 token/s
```

Mixed batches confirmed in scheduler logs (`Prefill batch ... #running-req: 36`).

8/8 outputs on a separate hand-crafted long+short prompt mix exactly match
a `--enable-dp-attention` (no mixed_chunk) baseline run, indicating
correct cross-DP token routing under MIXED.

Local run had to substitute `deepseek-coder-v2-lite` (cached) for
`lmsys/sglang-ci-dsv3-test` (gated on this box) and use
`--attention-backend triton --disable-cuda-graph` (an outdated local
`sgl-kernel` install passes an unsupported `out=` kwarg to FA3). The
registered CI test does not include those overrides — CI will run with
the canonical FA3 + cuda graph configuration.

Speed Tests and Profiling

N/A — bugfix on a previously crashing path.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

[gemini-code-assist]

Code Review

This pull request updates the batch overlap logic to include ForwardMode.MIXED when computing split sequence and token indices. It also adds a new test case, TestDPAttentionMixedChunk, to verify DP attention with mixed chunking. The reviewer suggested using the in operator for enum comparisons to improve readability.

[ch-wan]

/tag-and-rerun-ci