pxq_llama: PXA-native low-bit MoE quants (PXQ2/3/6 + PXQU) + fused Pascal/Volta kernels

[poisonxa16]

This is a fork of ik_llama.cpp that adds PXQ - a PXA-native low-bit quant family for MoE models, with fused CUDA kernels written for Pascal (sm_60) and Volta (sm_70). The goal was to put a real 35B MoE on the cheap 16 GB datacenter cards that modern quant kernels skip.

Results (all measured; repro harnesses in bench/):

• PXQ3: 14.7 GB, +1.1% perplexity vs the flagship quant (7.44 vs 7.36, wikitext-2, 200 chunks) - one 16 GB card
• PXQ2: 10.7 GB, fits an 11 GB 1080 Ti (+14% ppl)
• PXQU-16 (14.0 GB): a knapsack-mixed universal tier that runs full ub2048 prefill on a single 16 GB card - where ik IQ3_K (15.1 GB) OOMs at ub2048 single-card
• decode ~56 tok/s on a $100 P100 (flagship-quality PXQ3), ~99 on a V100, ~71 on a 1080 Ti; ~1950 prefill (V100)
• KLD vs bf16: PXQ6 0.056 / PXQ3 0.076 / PXQ2 0.205 (monotonic); SWE-bench Lite 19.6% (11/56, official docker grading); 0 refusals / 57 probes

Format sketch: expert tensors use a learned codebook + per-row E16 scales (a per-row fp16 anchor amortized over a 64-row panel + a 4-bit sub-scale per 16-elem block); backbone stays MXFP4. Every fast kernel is memcmp bit-exact against its baseline (bench/determinism-gates.md).

We also publish a negative result: protecting the attention backbone at q6 (like ik IQ_K) is a KLD wash at fixed size on Pascal and costs 3-5% decode, so we shipped the faster version - the override code is in the repo for anyone who wants the trade.

Built on ikawrakow's ik_llama.cpp - thank you for the base engine and the IQ_K line of work. Weights: https://huggingface.co/poisonxa/PXA-Fusion2-35B-GGUF

Feedback on the format or the kernels very welcome.