Add zero-shot benchmark results (ARC, HellaSwag, SciQ, PIQA) and full-split val ppl (#1)

TinyStories-24m: 24.59M BPE GPT trained from scratch on TinyStories (val ppl 8.76, coherent)