fix(cuda): pin speculative expert kernel family

[ZacharyZcR]

Summary

• extend SPEC_PIN to the synchronous CUDA grouped-expert path so speculative verification cannot switch into row-count-gated W4A4 or W4A16 Tensor Core kernels
• keep the shared-expert W4A16 path off during a pinned speculation window
• add an optional DLL symbol; older CUDA DLLs fail closed to the CPU fallback when pinning is required

This addresses the uncovered CUDA half of #689: CPU dispatch was already pinned, while CUDA grouped experts could still change numeric families with verify-row depth. The existing non-speculative path is unchanged.

Verification

• make -C c test-c
• make -C c glm
• make -C c glm CUDA_DLL=1 CC=x86_64-w64-mingw32-gcc
• git diff --check

test_backend_cuda now checks that a pinned group remains byte-identical to the small-batch grouped path even when W4A4 or W4A16 Tensor Core dispatch is enabled. This host has no NVIDIA GPU, so H100/H200 deep-draft token identity and acceptance-rate reproduction remains a hardware validation gate rather than a claimed result.