Assorted evaltune improvements

[TheRealGioviok]

This patch brings an overall training time reduction by a factor of 3 by introducing:

• Faster overall per epoch tuning (about 25% speedup)
• Much more efficient tuning schedule (1000 -> 450 epochs needed for convergence)
• Reintroduce Value::sum for the loss reduction operation. This is kind of an hack for now, as it can only be used once per graph cleanup (which is fine for now, since this patch uses it only for the loss reduction).

The first speedup is achieved by avoiding to populate the autograd tree with 0 sum and 0 products.
The tuning schedule is more reliable (observed lower rate of infamous king safety flukes, where ks values seem to just shoot towards high negatives for eg and then die gradient-wise), produces interpretable material values (yay!) and might even gain a couple of elo points.

================================================

Passed STC nonregression:

Test  | evaltuneoptim
Elo   | 2.55 +- 2.76 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [-3.00, 0.00]
Games | N: 20402 W: 5303 L: 5153 D: 9946
Penta | [203, 2494, 4700, 2558, 246]

https://ob.cwchess.org/test/1345/

LTC nonregression pending:

Test  | evaltuneoptim
Elo   | 2.03 +- 2.42 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [-3.00, 0.00]
Games | N: 22424 W: 5531 L: 5400 D: 11493
Penta | [118, 2652, 5538, 2789, 115]

https://ob.cwchess.org/test/1346/

Bench: 17090943