Restore Julia-native server (replace Python/FastAPI with Flux.jl + HTTP.jl)

Fix: accept null optional params, add exception handlers for chatroom compat

Match RandyGPT format: open CORS, no role chunk, plain dict response

Fix streaming: truly incremental SSE with initial role chunk

Fix default temp=0.8, top_k=40 for better output quality

Switch to Python/FastAPI server (RandyGPT pattern)

Remove TiedDense, use direct weight-tied output projection

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

Fix inference: vocab mismatch, sampling, and performance

Fix loadmodel for TiedDense weight sharing

Add HF model download at startup + Downloads dep

Add JuliaFluxGPT inference server (Flux.jl LLaMA-style GPT)

initial commit