`AVM`: Optimizations

[iakovenkos]

1. Avoid creating zero polynomial of fixed circuit size in the prover before batching -- tweak compute_batched to compute the linear combination in place, i.e. initialize with the first entity, and swap the accumulator if a poly of bigger size is encountered.
2. Try multi-threading the above using some entities as acccumulators
3. Hash shifted polys only once. Compute batching challenges for unshifted instead (maybe it has to be AllValues, not just a vector), compute batched_shifted using corresponding short challenges, compute batched unshifted as $\sum_{\text{non-shiftable}} \rho_i * P_i + \text{batched-shifted}$

[iakovenkos]

@fcarreiro we discussed 1. before the break, @federicobarbacovi will tackle it and couple of other issues as he's looking into AVM at the moment

[fcarreiro]

Nice thanks!, some comments of things I already tried

1. Note that right now we have a column precomputed_clk which goes from 0 to 2^21 (ie circuit size). This will always be the biggest one. You can get the index using Column::precomputed_clk. Because of this, I was wondering if in the AVM we need to commit to/open the (124) precomputed columns, since the VK is fixed. Avoiding it completely would be awesome.
2. I tried this and it was slower BUT i was creating new polys, if you can use existing entities then it might just work great.
3. no comments here :)

Reach out to me if you want to discuss avm particularities.