maximize the usage of S::BaseField to have smaller proof and better performance

[kunxian-xia]

The arithmetic between two base field elements b op b (where op can addition, subtraction, multiplication) is much faster than e op e. This fact also applies in e op b.

Currently, PR #17 uses degree 2 extension field of goldilocks ScalarExt2 to store the R1CS matrices which implies the phase 1 sumcheck is done over the extension field. This is not optimal.