Add multi-LoRA MoE inference support

[kiddyboots216]

Summary

• Adds multi-adapter LoRA inference for Mixture-of-Experts models, enabling concurrent serving of multiple MoE LoRA adapters in a single batch
• Implements full MoE LoRA forward pass in the unquantized Triton backend using CSGMV kernels for gate_up_proj and down_proj shrink/expand
• Adds per-expert LoRA weight buffers (A/B matrices + presence masks) in the memory pool, per-token adapter index tracking in the chunked SGMV backend, and CUDA graph bypass when MoE LoRA adapters are active

Changes

File	Description
lora/moe.py	New module — MoE LoRA constants, weight normalization, chunked compound segment builder
lora/lora.py	Load and normalize MoE weights in LoRA adapters
lora/mem_pool.py	Allocate/load/evict per-expert LoRA weight buffers
lora/lora_manager.py	Wire MoE modules, set MoE LoRA info per batch
lora/utils.py	Per-token adapter index expansion, MoE target module normalization
lora/backend/chunked_backend.py	Token-level weight indices for CSGMV backend
layers/moe/fused_moe_triton/layer.py	MoE LoRA hooks on FusedMoE layer
layers/quantization/unquant.py	Full MoE LoRA forward pass (gate_up + down with validation)
model_executor/cuda_graph_runner.py	Disable CUDA graph for MoE LoRA batches
model_executor/piecewise_cuda_graph_runner.py	Same for piecewise runner
test/registered/lora/test_moe_lora_utils.py	Unit tests for MoE LoRA utility functions

Limitations

• Only supports unquantized Triton MoE backend (not triton_kernels, FlashInfer CUTLASS, or TRT-LLM)
• Requires --lora-backend csgmv
• TP=1 and EP=1 only
• CUDA graph is bypassed when MoE LoRA adapters are active
• Does not support fused shared experts or no_combine=True