CUDA build fails on every branch: candle-kernels redefines __hmax_nan/__hmin_nan under CUDA 12.6 #86

build-cuda-wheel.yml has failed on every run since 2026-10-04, develop and main included (latest: run 38780125426 on #85). The CUDA build claim is therefore unproven on every branch.

Cause, from the log: candle-kernels 0.11.0 src/compatibility.cuh defines __hmax_nan/__hmin_nan under #if __CUDA_ARCH__ < 800, and CUDA 12.6's cuda_fp16.hpp (lines 3288/3305) now defines them for every arch. With CUDA_COMPUTE_CAP=75:

src/compatibility.cuh(11): error: function "__hmax_nan(__half, __half)" has already been defined
2 errors detected in the compilation of "src/affine.cu".

Done when

• The CUDA build is green with compute capability 75 (vendored candle-kernels with a patch guarding the definitions on the toolkit version, or a toolkit pin that is recorded with its reason).
• The guard is nullified once to show the build goes red again.