perf: frontend.Compile with PLONK became too slow

[gbotrel]

This #153 fixed an issue where we would split linear expression from R1C into duplicate plonk constraints. But the perf impact is non-negligible.

Compilation, on some circuits (the ones with large linear expressions like std/algebra/scalar mul or std/groth16) CPU time will be spend mostly (95% +) in:
split -> cs.coeffID(...) divideLinearExpression computeGCD and big.Int mallocs...