cuda: sanitize non-finite router scores

[exochard]

Problem

The CUDA router computes probabilities and scores used by fixed-size top-k selection. A NaN or Inf in a router logit or bias can poison comparisons and leave an invalid expert selection, which later becomes an unsafe expert-tensor index.

Change

• Centralize router probability conversion and clamp non-finite results to zero.
• Treat non-finite router bias entries as zero during score comparison.
• Apply the same guards to scalar, parallel, and warp top-k kernels.

This keeps selection deterministic and prevents malformed numeric input from becoming an invalid routed-expert address. It does not change finite-input behavior.

Validation

Builds completed with no compiler warnings:

make -B ds4 CUDA_HOME=/opt/cuda CUDA_ARCH=sm_86
make -B ds4 CUDA_HOME=/opt/cuda CUDA_ARCH=sm_89

git diff --check is clean. Target hardware is an RTX 4060 Ti 16 GiB (sm_89); the guard is independent of model quantization and applies to the CUDA DeepSeek router.