Refactor: final exponentiation in pairings

[yelhousni]

@feltroidprime was porting the Karabina optimisations in our final exp to Garaga and encountered a problem with: https://github.com/ConsenSys/gnark-crypto/blob/f93a56c714c4e6266429cac111a004e9eec7daa0/ecc/bls12-381/internal/fptower/e12.go#L227
When computing e(P,Q) * e(-P,Q) the decompression fails (in the corresponding Garaga code) at: https://github.com/ConsenSys/gnark-crypto/blob/f93a56c714c4e6266429cac111a004e9eec7daa0/ecc/bls12-381/internal/fptower/e12.go#L258
This is because g2 and g3 are both zero in that test case. In gnark-crypto this test does not fail because the case g2==g3==0 is handled "implicitly". Since we take the convention of 1/0 = 0 the returned value in this case is 1, which is what is expected as per the proof of Theorem 3.1 of Karabina's paper.

This is said, in this PR I make a suggestion to:

• "Explicitly" handle the g2==g3==0 case, and
• Return 1 when the result of the easy part is 1 — which is what happens in similar cases to e(P,Q) * e(-P,Q). This saves the hard part computation since 1^d==1.