Add Picolo-BananaMind using acc_norm for all benchmarks
Replaces closed PR #226. All submitted benchmark values are acc_norm (percent): ARC Easy 25.63; ARC Challenge 27.82; HellaSwag 23.70; PIQA 50.71; ArithMark 3 28.40. Full zero-shot test sets using local lm_eval; unmodified upstream ArithMark script, 1000 questions. Results and both raw/normalized metrics: https://huggingface.co/BananaMind/Picolo-BananaMind/tree/main/evaluation . Exactly 100 unique trainable parameters, causal self-attention transformer, safetensors and trust_remote_code. Trained on exactly 10M streamed FineWeb-Edu nibble tokens (two tokens per UTF-8 byte, plus document EOS; not ordinary BPE-equivalent). Training context 256; lm_eval context 8192; ArithMark context 1024. Novelty/chance-level experiment, not a reasoning claim. No ArithMark-2 result.