Allow negative humanSLChosenMovePiklLambda to prefer worse moves

[ChinChangYang]

The pikl lambda is an inverse temperature on KataGo's own utility applied to the human SL policy at final move selection, but it was clamped to non-negative values, so it could only ever shift play toward the moves KataGo likes best. Extend it to negative values so the same knob runs the other way:

small positive (0.01) - always the best move KataGo searched
large positive (1e9) - purely the human SL policy
large negative (-1e9) - purely the human SL policy, same as +1e9
small negative (-0.01) - always the WORST move KataGo searched

The formula already did this for a negative lambda, but subtracting the maximum utility makes the exponent positive there, so exp overflowed to +inf for any |lambda| below roughly 0.005 at realistic utility spreads. That produced 0.0 * inf and inf/inf NaNs which reached playSelectionValues and tripped testAssert(maxValue > 0.0), i.e. a fatal exit mid-game.

So subtract the minimum utility instead when the lambda is negative, which keeps the exponent <= 0 and exp within (0,1] no matter how small the lambda gets. The subtracted utility is only a constant factor and cancels in the renormalization, so nothing else changes. Use std::signbit for the test so that -0.0 counts as negative, and map a zero utility difference to a multiplier of 1 so that -0.0 gives the limit (only the worst move keeps weight) rather than 0/0. Moves that cannot be chosen anyways do not set the reference, since otherwise every other move could underflow to zero at once - pass is one of those when humanSLChosenMoveIgnorePass is leaving passing up to KataGo.

The non-negative path is left exactly as it was, including the existing quirk that a lambda of exactly 0.0 disables the blend rather than meaning "always the best move".

Claude-Session: https://claude.ai/code/session_01R53aXQZ6fprZRKsV7on3pu