feat(infiniop): add lightning_attention op (CPU + NVIDIA) for MiniMax

[y258dd]

NVIDIA RTX 4090 / sm_89 Validation (2026-09-19)

Validated with:

• InfiniCore feat/minimax-lightning-attention at 7bcaa623
• InfiniLM feat/minimax-model-support at e23eba79
• RTX 4090, compute capability sm_89, driver 570.124.06, CUDA 12.8 (V12.8.61)
• PyTorch 2.6.0a0+ecf3bae40a.nv25.01 with CUDA 12.8, torch.cuda.is_available() == True
• Python 3.12.3, g++ 13.3.0, xmake 3.1.1+20260919

Results

Test	Device	Result	Error
MiniMax lightning attention operator	NVIDIA	Passed	decode: out 4.768e-07, state 2.384e-07; prefill: out 1.907e-06, state 4.768e-07
MiniMax smoke prefill/decode	NVIDIA	Passed	finite logits; state continuity 3.127381e-06
MiniMax vs HuggingFace, 1 expert, FP32	NVIDIA	Passed	prefill 5.257875e-04; decode 2.117679e-04
MiniMax vs HuggingFace, 4 experts, top-2, FP16	NVIDIA	Passed	prefill 7.324219e-04; decode 3.662109e-04

All tests are below the 1e-2 model-level tolerance. The 4-expert top-2 result exercises the CUDA Fused MoE path rather than the dense fallback.

Fixes Included

• CUDA Lightning Attention now supports FP16 and BF16 in addition to FP32.
• MiniMax linear-attention state uses the model dtype instead of always allocating FP32.
• Sampling offsets are copied to host before host-side indexing on CUDA.
• HuggingFace MiniMax MoE weights are remapped to InfiniLM expert keys.
• Lightning-attention GPU tests use disjoint initial/final state rows to avoid cross-request races.

Server Logs

• /data/minimax-final-gpu-op.log
• /data/minimax-final-gpu-smoke.log
• /data/minimax-final-gpu-vs-hf.log
• /data/minimax-gpu-test-moe4-top2-fp16-final.log