Fix default temp=0.8, top_k=20, fix fallback tokenizer to 38 chars

Switch to Python/FastAPI server (RandyGPT pattern)

Aggressive sampling: temp=0.1, top_k=8, rep_penalty=1.3

Fix inference: top-k sampling, repetition penalty, prompt encoding, vocab download

Upgrade to juliaflux v1: Flux.jl GPT-2 architecture with HF model download

Upload checkpoints/best_model.json with huggingface_hub

Upload server.jl with huggingface_hub

Upload checkpoint.jl with huggingface_hub

Upload model.jl with huggingface_hub

Upload Project.toml with huggingface_hub

Upload Dockerfile with huggingface_hub

Upload README.md with huggingface_hub

initial commit