Function softmax_kernel_source 

pub fn softmax_kernel_source(k: usize, p: usize) -> String

Prepend the KK / PP macros so the NVRTC compile is a pure compile_ptx, matching sphere_gpu / arrow_schur_nvrtc.

Also prepend an INFINITY definition: the kernel seeds its softmax max reduction with -INFINITY, but NVRTC does NOT predefine INFINITY (it is a <math.h> macro, not a CUDA builtin), so without this the whole module fails to compile and the SAE row-jet path silently falls back to the CPU (same genus as the M_PI NVRTC fix). __longlong_as_double is an always-available NVRTC builtin needing no header.