GPU-unlock kernel module doubles BF16 instruction throughput on NVIDIA 3000 series

This Linux kernel module unlocks full FP/BF16 instruction issue rates on GA102-GA106 cards by writing speed selectors during pre-driver timing. It avoids driver patching by using a kmod approach, increasing BF16 performance from 87.50 TFLOPS to 157.41 TFLOPS and TF32 from 42.58 to 82.61 TFLOPS on a 3090 Ti.