auto-wit-jit-gen performance ideas

[chriseth]

• use MaybeUninit for the return values for a submachine call
• if the WitgenFunctionParams are not passed via registers, maybe change it to actual individual function parameters.
• remove the known field from CompactData in some cases: If we use the jit to run a block machine, we know the exact shape and can derive it if it is required.
• perform stores early to improve optimizer performance and reduce stack pressure
• sort stores by row and column id
• the fixed lookup machine allocates a BitVec (it contains a Vec) - can we avoid that?