This PR adds text-only support for Qwen3.8-Flash-Next (qwen4_exp) to FreeToken, serving 36 GDN and 12 QSA layers on 4 hyper-connection residual streams.

The model uses a PLE n-gram embedding on a 47.7 GiB pinned-host table, 512 NVFP4/block-fp8 routed experts with a top-10 router, and a gated shared expert. Triton kernels implement QSA sparse attention, PLE gather, MoE routing, and radix block top-k, with hybrid prefill chunks aligned to page boundaries.