Rescoring doesn't seem to work with NNUE based eval.

[dpldgr]

Rescoring 200M depth 2 positions to depth 10 after confirming a NNUE file was actively being used resulted in similiar performance to HCE. In theory, rescoring to higher depths using the same network should result in more accurate rather than substantially worse evaluation, and it has the benefit of being faster than simply using the higher depth to generate the data in a single pass using the generate_training_data command.

See discussion here:

https://discord.com/channels/779317816897699850/966610323987660830/1222815835119288330

For an idea of the potential increase in data gen speed see here:

"As a rough calculation, I was getting ~1600 sfens/sec with d2 generate_training_data, then ~1800 sfens/sec with d10 transform rescore, so if we go the full gen 200M d2 and rescore to d10 each generation, that still works out at 847 sfens/sec rescored, faster than I was getting with d6 generate_training_data (~650 sfens/sec)!!"