Remove some more CPU cycles

for most loops iterating over i until i == len, and reading/writing to the same memory segment, both i and len are scaled from the beginning by the corresponding segment value (and unscaled while leaving the loop)
packs all 17 0 writes when creating a new branch node into a single write
removes the txn bloom segment reset post execution (we only process 1 txn per proof now)

Shaves off between 0.6% and 1.8% of CPU cycles for a set of arbitrary L1 txns, on top of #1469 (which already removed ~7%/8%).

0xPolygonZero / plonky2