LinearOpTypeMap specializations for CUDA device targets.

Provides compile-time dispatch for (Cuda, TPrecision, TWeightQuant) combinations. A missing specialization for any combination is a hard compile error.