Positive policy decay

[Naphthalin]

This is a reimplementation of #1173 and #1288 without the child visit boost part, aimed at supplementing CPuct scaling. Setting either PolicyDecayFactor or PolicyDecayExponent to 0 effectively turns the policy decay off, so it should be tuning friendly.

[Naphthalin]

Some words of explanation:

• PUCT in the way Alphazero uses it treats policy as a multiplicative factor to the U term, persistent even at millions of nodes. The original reference meanwhile treats policy as an additional "delaying" term, decaying to 0 with more nodes.
• this approach (first version was a semi April fools joke in 2020) picks up the general idea of decaying the policy effect, but instead by affecting the multiplicative P% term, keeping the nature of the PUCT formula intact.
• it's crucial that policy decay (and generally any modification to PUCT) is done in a way that the U term still gets reduced after a visit, and slightly increased after a sibling is visited. Hence why "positive policy decay" is needed, otherwise search inconsistencies like Add visit-based policy temperature decay #1150 happen.
• what this does is effectively removing the "policy sums up to 100%" condition, resp. effectively increases cpuct especially in positions with multiple good moves in addition to cpuct scaling.
• mathematically, the formula converts policy into a logit, and adds a logarithmically growing term. With N->inf, all P% will grow towards 100%, though the initial P% value decides how much this growth is delayed.

I expect this to work best together with a reduced CPuctFactor, with PolicyTemperature, CPuct and FpuValue optimized for STC. A tuning could therefore happen in two steps; first the STC tune (~1k npm) with the 3 parameters, and then a LTC tune (>100k npm) with CPuctFactor, PolicyDecayExponent and PolicyDecayFactor.