Reduce the cost of a bus interaction

[georgwiese]

We currently commit to two extension field elements per interaction (for GL: 2 cols / ext; for BB/M31: 4 cols / ext).

We can reduce them in two ways:

1. One is the folded tuple. The reasons are here. We could remove it, but we would either have to disallow high-degree expressions and next references, or have a way to only materialize the folded tuple if necessary (for which we would need to inspect the expression, which we can't currently in PIL).
2. The accumulator: Here we could batch, and have one accumulator for several bus interactions (as they are all summed up globally anyway). Each additional interactions increase the degree of the constraints, but with a degree bound of 3, we could batch two interactions, so this would cut the cost in half. For this, we need a global view of all bus interactions though (this goes in the direction of the "capture constraints" feature Chris proposed).

So overall, it could be reduced by a factor of 4, but it is not simple to implement in PIL without additional language features.

But what we could do today:

1. Disallow expressions of degree > 1 and next references, as mentioned above, and remove the folded tuple commitment.
2. Add a multi_send function that generates one accumulator for two sends (for receives, we should only have one per machine to begin with). Use this in the ASM linker.