support kimi 2.5/6 lora (logprob diff exist)

[nanjiangwill]

Adds Kimi K2.5 / K2.6 LoRA support to the `sglang-miles` branch:

• MoE LoRA: `experts_shared_outer_loras` layout. `gate_up_proj_moe` has shared A (`(slot, 1, 2r, hidden)`) + per-expert B; `down_proj_moe` has the inverse. `LoRAMemoryPool` initializes A and B independently per target so a shared Tensor on one side doesn't clobber the per-expert dict on the other.
• Virtual-experts kernel: `lora_b` as a list. `merged_experts_fused_moe_lora_add` accepts a length-2 `lora_b` for the gated case (one `2r` shrink, two `r` expands); `lora_moe_runners` passes `tuple(lora_b_stacked)` for it.
• Chunked `load_lora_adapter_from_tensors`. Per-rank serialization: `serialized_tensors: str` -> `serialized_named_tensors: List[Union[str, bytes]]` so each TP rank deserializes its own producer's CUDA-IPC bucket (matches `update_weights_from_tensor`). `Engine` and `BaseTpWorker` plumbed accordingly; `TokenizerControlMixin` relaxes `dp_size == 1` to also allow `enable_dp_attention`.
• DP-attention safety. Token→lora mapping pads `req_to_lora` with a `-1` sentinel so foreign tokens past `seg_indptr[-1]` don't `searchsorted` past the end / read uninitialised memory.
• Quant post-process skips LoRA wrappers. `restore_weights_before_load` / `process_weights_after_loading` loops in `ModelRunner` skip `BaseLayerWithLoRA`; the inner base layer is yielded separately by `named_modules` and gets processed there.
• MLA fused-A cache persistence. `DeepseekV2WeightLoaderMixin` persists `cached_a_proj` across `do_load_weights` calls so chunked updates that split `q_a_proj` / `kv_a_proj_with_mqa` across chunks still produce the fused `fused_qkv_a_proj_with_mqa` once both halves arrive. (Reverted on `sglang-miles` in Revert deepseek_weight_loader.py changes from #25141 #28561 — sender-side pairing in Pair q_a_proj / kv_a_proj_with_mqa in the bridge weight iterator radixark/miles#1361 is the cleaner fix.)
• Base-layer dim derivation. `BaseLoRABackend` derives MoE intermediate buffer dims from the base FusedMoE rather than quant-specific tensors so it works for any scheme (FP, WNA16, Marlin-packed, …).

`logprob diff exist` in the title flags a residual `train/train_rollout_logprob_abs_diff` drift seen on K2.5 RL runs;

---

Decomposition for `main`: #28562, #28564, #28566. The MoE post-process LoRA-wrapper skip is blocked on #28565 (upstream prereq). The MLA fused-A cache change is being reverted on `sglang-miles` via #28561.

[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!

[yushengsu-thu]

these modification should be sent to sglang/main as well or it will becomse useless after we upgrade the sgl

[yushengsu-thu]

the modification are almost lora related only ‎python/sglang/srt/models/deepseek_common/deepseek_weight_loader.py‎ is related to kimi2.5 base.

We can consider if revert the kimi2.5 base related parts. To unblock kimi full-param and lora training support so I merge them first.