examples: change GPT-2 KV cache to fp16 to take advantage of tensor cores

[bssrdf]

This PR changed default KV cache type to FP16 to use tensor cores if available. This brings the inference performance much closer to llama.cpp.

[ggerganov]

This is a bit irrelevant for the purpose of this example, because it is not meant to deliver optimal performance, but rather to demonstrate ggml usage. I think we can simply add a comment here that either GGML_TYPE_F16 or GGML_TYPE_F32 can be used.

[bssrdf]

@ggerganov, fair enough! I have reverted back to FP32 and added a comment. Thanks for reviewing.