add llama 3 support to llm.c

[karpathy]

This branch starts with a copy paste of train_gpt2.cu and test_gpt2.cu, but these two files (and other files) will change to incorporate Llama 3.1 support, before merging back to master.