Make whisperfile 12% faster in GPU mode

whisper_process_logits() was computing soft max on its own many times so
I changed it to call my vectorized expf() function ggml_vec_soft_max_f32
which was upstreamed to llama.cpp a few months ago. Since this is pretty
much the only CPU operation that happens in GPU mode, it has a very huge
impact on performance compared to llama's large language model inference