chore: minimize usage of get_row in inverse computation

[ledwards2225]

Using get_row in hot loops is known to be a bad plan since it amounts to a copy. When I implemented log derivative table lookups, I carelessly used the method in logderiv_library.hpp which computes the inverses needed in the log deriv relation. That method uses get_row in a loop over the entire execution trace which amounts to copying the entire set of prover polynomials one row at a time.

A half-measure fix is to still use get_row, but only where we really need to use it (i.e. where the gate in question is active). this is what I did for the databus log-deriv lookup relation. This is still less efficient than it could be because we're copying an entire row when we only need a small subset of the data. This too could be avoided but at the cost of some icky duplication so for now I'm leaving it as is with reference to a bberg issue.

Results in a ~0.5s reduction in ClientIvc bench, essentially entirely in the PG prep round (where the lookup inverses are computed):

Before:
ClientIVCBench/Full/6 15131 ms 10319 ms
ProtoGalaxyProver_::preparation_round(t) 3569 52.15%

After:
ClientIVCBench/Full/6 14582 ms 9853 ms
ProtoGalaxyProver_::preparation_round(t) 3055 48.72%