gumbel-mcts ships numba-accelerated MCTS library

This is the kind of quiet infra release that matters to anyone actually running search-heavy agents. The real win is less the algorithm name than the combination of validated parity, low-budget Gumbel search, and a compact codebase people can inspect and fork.

• –`puct.py` claims exact policy parity with a reference MCTS while delivering a multi-x speedup.
• –`gumbel_dense.py` is the more interesting part for MuZero-style setups where simulation budget is tight.
• –`gumbel_sparse.py` is a practical fit for large branching factors, which is why chess is called out explicitly.
• –MIT licensing and a pure Python/numba stack lower the barrier for experimentation, benchmarking, and integration into self-play loops.