[MB_KECCAK_PECTRA] keccak abi optims

[mabbamOG]

What ❔

• keccak abi now performs rotations only in unique precompile modes -> 30% more cycles, but better packed circuits
• keccak abi now delegates final round-constant xor to a mock 25th round precompile call, we throw away the non-iota computations
• added leaner perf.sh script to measure cycle count
• streamlined airbender imports to workspace
• changed some airbender imports to mb_dev branch, must be changed back!

Why ❔

we wanted a more compact keccak circuit, it was wasting 25% of constraints/vars on 75% of cycles, roughly. Now we waste 50% of constraints/vars on 25% of cycles. Removing that final waste i consider to be insane work for minimal benefit.

Is this a breaking change?

• Yes
• No, but use updated airbender branch

[mabbamOG]

also, i did perform measurements of the "one free limb" u8 chunking trick, as opposed to the current approach where we create all u8 variables and then do quadratic constraints for in/out wiring.

they cost the same, so for simplicity i kept the current method which generates all lookup u8 variables

@shamatar