[cherrypick from #36389] [LoRA] Let DP-attention LoRA run idle ranks and eager prefill (partial)

[yushengsu-thu]

Summary

Port two hunks of #36389 to the sglang-miles branch: its managers/scheduler_components/dp_attn.py and lora/utils.py changes. Without them, a LoRA server with --enable-dp-attention dies on its first batch:

• AttributeError: 'EagerRunner' object has no attribute 'can_replay_locally' in _local_prefill_cuda_graph_vote. With LoRA the prefill runner is the EagerRunner fallback, and the vote now returns False for it.
• ValueError: Unsupported forward mode: 4 in get_batch_token_counts, from MoE-LoRA batch preparation. DP ranks without requests run IDLE batches, which get_batch_token_counts and generate_sequence_lengths now accept.

Port notes

• This is a partial cherry-pick. The changed lines are identical to [sglang][lora] Support DP attention in LoRA backends #36389's changes in these two files.
• The rest of [sglang][lora] Support DP attention in LoRA backends #36389 is not included: the DP-local / TP-global LoRA routing and its pinned-adapter and dp_size == tp_size requirements. It overlaps with sglang-miles' own DP-attention MoE-LoRA path and conflicts in 7 files.
• Independent of [cherrypick from #39379] [LoRA] Size dense row/column-parallel LoRA buffers from the base linear's real shard #41506 in code. The GPU test below runs both, because under --moe-dense-tp-size 1 the server needs [cherrypick from #39379] [LoRA] Size dense row/column-parallel LoRA buffers from the base linear's real shard #41506 to finish CUDA graph capture first.

Testing

• Unit tests (8×H200 devbox): test/registered/unit/lora/, test/registered/unit/managers/scheduler_components/test_dp_attn.py, test/registered/unit/managers/test_scheduler_hicache_events.py and test/registered/unit/managers/test_scheduler_chunked_req_gate.py have 177 passed with and without this PR. The 4 failures are identical.
• GPU: DeepSeek-V2-Lite on 8×H200 with --tp 8 --dp 8 --enable-dp-attention --ep 8 --moe-dense-tp-size 1 --enable-dp-lm-head, triton MoE and LoRA backends, and a LoRA on the layer-0 dense MLP, on top of [cherrypick from #39379] [LoRA] Size dense row/column-parallel LoRA buffers from the base linear's real shard #41506:
  • Without this PR, the first batch fails with the EagerRunner error under the default prefill graph settings, and with the IDLE error under --cuda-graph-backend-prefill disabled.
  • With this PR, both settings serve 32 concurrent mixed base/LoRA requests with no scheduler exceptions. LoRA parity is reported in [cherrypick from #39379] [LoRA] Size dense row/column-parallel LoRA buffers from the base linear's real shard #41506.
• pre-commit run on the changed files passes.

---

CI States

Latest PR Test (Base): ❌ Run #36483117678
Latest PR Test (Extra): ❌ Run #36483116907
Latest PR Test (AMD ROCm 10): ❌ Run #36483117405