perf: improve fft domain memory footprint

[gbotrel]

This PR removes precompute of CosetTableReversed and CosetTableInvReversed in fft.Domain.
In practice these can be accessed from CosetTable and CosetTableInv .

New version may result in slightly slower FFT on small domains.

However, for large domains (think large plonk circuit), precomputing these and storing them in RAM contribute to a high memory load, hence the PR.

[yelhousni]

This makes sense for large zkEVM circuits indeed. What is the exact slowdown on small circuits, say a 2-chain Groth16 verifier?

[gbotrel]

Benchmarked groth16_bls12377 , virtually no delta on the M1 👍

[gbotrel]

1% slower for a larger circuit.