[LoRA] Sentinel-pad token->lora mapping for DP-attention foreign tokens

[nanjiangwill]

Summary

Under DP-attention, the gathered forward batch contains both local and foreign tokens, but `seg_indptr` / `weight_indices` cover only the local requests. `_compute_moe_lora_info` in `backend/base_backend.py` currently does one of two wrong things at the foreign positions:

• CUDA-kernel path: `token_lora_mapping` is `torch.empty(...)`, so foreign-position slots stay uninitialized when the kernel only writes to positions within segments. The MoE LoRA hook then dispatches on garbage.
• Python fallback path: `torch.searchsorted(seg_indptr, ..., right=True) - 1` returns `num_segments` for foreign positions, which makes `torch.index_select(weight_indices, 0, req_indices, ...)` index one past the end of `weight_indices`.

This PR fixes both paths with the same -1-sentinel convention:

1. Pre-fill `token_lora_mapping` with `-1` before either path runs, so un-segmented slots stay `-1` after the kernel returns.
2. In the Python fallback, `torch.cat` a `-1` entry onto `weight_indices` and use that padded tensor for the `index_select`, so the same `-1` lands at the foreign positions.

Foreign-token LoRA outputs are discarded by the DP-attention scatter anyway, so mapping them to `-1` ("no LoRA") is the correct local behaviour — it just stops the kernel from acting on uninitialised data / out-of-bounds reads.

Context

Carved out of #25141. The sglang-miles version of this fix lives in a helper called `_compute_token_lora_mapping` in `lora_moe_runners.py`; on `main` the equivalent work is done in `_compute_moe_lora_info` here, so this PR ports the same idea to the on-main code path.

Test plan

• Existing non-DP-attention LoRA tests — behaviour unchanged (the pre-fill is a no-op when the kernel/fallback overwrite every slot, which is the normal case).
• DP-attention + MoE LoRA: confirm the kernel and Python paths both place `-1` at foreign positions.
• Empty / no-segment edge cases — covered by the simplified `num_tokens == 0 or not has_segments` return-early (the prior explicit `fill_(-1)` for the no-segments case is subsumed by the earlier pre-fill).

---

CI States

Latest PR Test (Base): ❌ Run #27722492929
Latest PR Test (Extra): ❌ Run #27851913609

[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!

[nanjiangwill]

not needed anymore