add learn-phase thread benchmark

bench: __main__ guard for spawn

parallel self-play (--workers), conversion experiments, whitelist

fix: disable gradio SSR (node server crash)

tinychess: self-play research substrate (phase 1-4)