perf,memory: lighter plonk ProvingKey (no trace)

[gbotrel]

For a 500k constraint circuit:

BenchmarkSetup/bn254-10          357047889      204335417      -42.77%
BenchmarkSetup/bls12_377-10      551201334      304352802      -44.78%
BenchmarkSetup/bw6_761-10        2798877250     1690444708     -39.60%
BenchmarkProver/bn254-10         731693458      770120688      +5.25%
BenchmarkProver/bls12_377-10     964496188      1016455625     +5.39%
BenchmarkProver/bw6_761-10       4322171083     4459419875     +3.18%

This PR removes the Trace from the ProvingKey; the Prover computes it at runtime, it's not super expensive but avoids doing some Clone() (memallocs).
Setup time is also faster because we now use the kzg in Lagrange form to commit, and don't need to compute the large fft.Domain.

Memory wise on bw6 for example, we go in the Prover from:

begin Alloc = 246 MiB   TotalAlloc = 1051 MiB   Sys = 394 MiB   NumGC = 22
end Alloc = 388 MiB     TotalAlloc = 1193 MiB   Sys = 446 MiB   NumGC = 22

to

begin Alloc = 150 MiB   TotalAlloc = 739 MiB    Sys = 246 MiB   NumGC = 21
end Alloc = 285 MiB     TotalAlloc = 957 MiB    Sys = 344 MiB   NumGC = 23

so a practical ~22% less memory allocated .

• in practice saves a lot of time on ProvingKey deserialization.

[ThomasPiellard]

Normally domain.Cardinality should match size := ecc.NextPowerOfTwo(sizeSystem), don't know if we should add this sanity check...

[ivokub]

Looks good!