v2 loader: VOCAB 8192 -> 12288 (#8)

v2 config: vocab 12288, 56,902,144 params (#7)

v2 card: 12288-vocab retrain, 56,902,144 params, val 1.3837, measured evals (#6)

v2: 12288-vocab BPE tokenizer (was 8192) (#5)

v2: retrain with 12288-vocab BPE (was 8192), 56,902,144 params, best val 1.3837 (#4)

Add weights (54,804,992 params, 99 tensors, float32)

Add BPE-8192 tokenizer

Add self-contained loader + TinyStoriesGPT class

Add architecture config (D512/L16/H8/FFN2048/SEQ512/VOCAB8192, weight-tied)

Add model card: 54.8M from-scratch TinyStories GPT, verified params + held-out PPL 5.24

initial commit