KataGo analysis engine requires nnMaxBatchSize in config (benchmark says to remove it)

When wiring KataGo (v1.17.1, JSON Analysis Engine) behind a WebSocket bridge for a browser Go app on macOS/Metal, two non-obvious things cost time:

1. katago benchmark prints: WARNING: Your nnMaxBatchSize is hardcoded, recommend deleting it. But if you actually delete nnMaxBatchSize from analysis.cfg, katago analysis aborts at startup with: uncaught exception IOError: Could not find key nnMaxBatchSize in config file. The ANALYSIS engine treats it as required even though the benchmark recommends omitting it. Keep the line (e.g. nnMaxBatchSize = 16).

2. An analyzeTurns query returns ONE response per turn (each tagged turnNumber), all sharing the same query id. A bridge that deletes its id->socket routing entry after the first response will drop the rest. Size the routing entry to analyzeTurns.length and only free it after all responses (or on error/close).

Also: reportAnalysisWinratesAs = BLACK makes winrate/scoreLead Black-relative; ownership is a flat size*size array, Black-positive, row-major from top-left (index=y*size+x). Empty 19x19 board top moves come back as the four 4-4 points (Q16/Q4/D16/D4), a good sanity check. b18c384nbt on Apple M4 Pro Metal benchmarks ~115 visits/s at numSearchThreads=8.