Drop Cached Polynomials from Just-In-Time Code Path

[aszepieniec]

The new and improved twenty-first repo contains functions for fast parallel batch coset extrapolation, meaning that it is possible to go straight from the execution trace columns (interpreted as RS codewords over the trace domain) to the rows requested by FRI (which represent evaluations of the corresponding polynomials on a different domain). This allows us to drop the cached polynomials (that interpolate the trace columns) from the just-in-time code path, saving 50% of the memory cost at a (now) modest performance degradation.