feat: custom_hf without CUDA falls back to offline baseline

fix(custom_hf): prefer dtype over deprecated torch_dtype in from_pretrained

perf(ui): fewer Gradio steps + live STEPS_PER_EPISODE, tighter token cap

perf(space-ui): faster Gradio custom_hf — UI candidates=1, cap tokens, GenerationConfig

Finalize eval-friendly defaults: offline baseline, deterministic API reset, docs cleanup

docs+router: HF Space secret names; auto includes custom_hf only when HF_TOKEN set

fix(router): exclude custom_hf from auto chain so Space auto stays fast

perf(ui): default mock provider; auto tries NIM/OpenRouter before HF; cap HF router timeout 90s

feat: inference router (HF/NIM/OpenRouter/mock); README aligned with judge rubric; chart axis labels