KataHex 20240812

This is a new version of KataHex (Katago for Hex game)
Using newer model structure (b18c384nbt) and with much more training, the new version becomes much stronger.

Running selfplay on large boards is too slow(~O(n^4) to O(n^7)). So I started training from a smaller board 15x15 (the 2022 version started from 13x13). At the final stage I continue training the model on larger boards, called "Transfer learning".
This version was trained on 15x15 board using 2*RTX4090 for 2 month, ~40M selfplay games.
Then on 19x19 for 15 days, ~2.6M selfplay games.
Finally 27x27 for 3 days, 90k selfplay games.

Old version: https://github.com/hzyhhzy/KataGo/releases/tag/20240406 (Using b40c256 net trained in 2022)
New vs Old, 500 games with balanced one-move openings, 100 visits per move.
13x13 board: +56 elo (290:210)
14x14 board: +273 elo (414:86)
15x15 board: +268 elo (412:88)
19x19 board: +263 elo (410:90)
25x25 board: +466 elo (468:32)
33x33 board: +511 elo (475:25)
Testing on 11x11 board is meaningless because it was almost solved even with any one-move openings.
The largest board which KataHex can play well is around 33x33 ~ 41x41.

There are two packages.
OpenCL only: Compatible for all GPUs. But slower than TensorRT on Nvidia GPUs
TensorRT+Cuda+OpenCL: For Nvidia GPUs only. Include three types of engines. TensorRT is the fastest but only available for Nvidia GPUs. This package include many DLLs of TensorRT and Cuda, so this package is much bigger.