Switched NNUE arch to bucketed

[liamt19]

These are two major changes wrapped into one patch:

• Replaced the "Simple768" architecture with a (768x4 -> 1536)x2 -> 8 network.
• Added Finny tables, which speed up accumulator refreshes

Almost certainly worse at STC because of the significant slowdown (~13%) from the king buckets:

Elo   | -6.25 +- 7.59 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | -1.34 (-2.94, 2.94) [0.00, 3.00]
Games | N: 4500 W: 1219 L: 1300 D: 1981
Penta | [62, 592, 1025, 507, 64]
http://somelizard.pythonanywhere.com/test/607/

But does quite well at longer time controls:

Elo   | 13.85 +- 8.88 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=32MB
LLR   | 2.94 (-2.94, 2.94) [0.00, 5.00]
Games | N: 3036 W: 845 L: 724 D: 1467
Penta | [11, 315, 762, 402, 28]
http://somelizard.pythonanywhere.com/test/608/