Pairing: remove preCompute and go routines in MillerLoop

[gbotrel]

@yelhousni following your recent awesome PR (#26) did some experiments and benchmarks on various architectures (x64), that confirms that for a small number of pair of points in the Pairing function, there is a severe penalty (specifically on amd platforms).

In this PR, I just removed the go routines, the multiple [68]lineEvaluation (replace by a single lineEvaluation object) and merged the preCompute loop within the line evaluation / result compute loop. Did on bn256 and bls381.

** If we want to deal with large number of pairs in Pairing, then we should probably revisit the preCompute strategy **

Here is where we're at on our reference benchmark (running on AWS z1d.3xlarge instance, with hyperthreading disabled) :

	mcl(ns/op)	gurvy(ns/op)
Pairing(bn256)	478244	489258
Pairing(bls381)	676513	707984

For the before/after removal of go routines and precompute:

on AMD Ryzen 3700X, bn256 and bls381:

benchmark                  old ns/op     new ns/op     delta
BenchmarkPairing-16        717847        429341        -40.19%
BenchmarkMillerLoop-16     330055        209102        -36.65%

benchmark                  old ns/op     new ns/op     delta
BenchmarkPairing-16        1156689       644036        -44.32%
BenchmarkMillerLoop-16     517074        308647        -40.31%

on Intel MBP, less conclusive but not too bad:

benchmark                 old ns/op     new ns/op     delta
BenchmarkPairing-8        561681        546137        -2.77%
BenchmarkMillerLoop-8     253963        263012        +3.56%

[yelhousni]

LGTM

[yelhousni]

@gbotrel In the last commit, I applied the same changes to bls377.