Initialization: GPU->host bandwidth could be halved

[poszu]

Currently, the code transfers 32 bytes of each label from GPU to the host. 32B are needed to search for the POPS VRF nonce because all 32B are compared with difficulty. Later on, only the most significant (big-endian) 16B are preserved in the POS data.

It's possible to optimize the bandwidth and send only the 16B of each label. The missing lower 16B could be generated on the CPU on the rare occasion (1/2^128) when the most significant bytes are not enough to decide.

Given a 16B label and 32B difficulty, there are 3 possible cases to consider:

• label > difficulty[0:16] -> NOT a nonce
• label == difficulty[0:16] -> POSSIBLY a nonce - generate lower 16 bytes of label
• label < difficulty[0:16] -> VALID nonce

The code would need to generate a label on the CPU only if the 16 bytes of the label are equal to the most significant 16B of the difficulty.