fix: clobbered bp issue in assembly on e2 impl

[gbotrel]

No API changes in this PR. Since Go1.16, go vet notices when BP register is clobbered with a non-zero frame size. Assembly code now doesn't use BP in those settings.

• updated to latest goff, with some minor refactoring on the asm code generation part.
• refactored E2.Mul and E2.Square assembly impl for bn256 and bls381 to be almost identical
• tried to not inline fp.Mul in assembly, to reduce code size. Behaves well on Intel, poorly on AMD, so reverted.
• this PR results in minor perf improvements:

bn256:

benchmark               old ns/op     new ns/op     delta
BenchmarkPairing-8      531315        510955        -3.83%
BenchmarkE2Mul-8        58.0          51.2          -11.76%
BenchmarkE2Square-8     37.1          35.6          -3.86%

bls381:

benchmark               old ns/op     new ns/op     delta
BenchmarkPairing-8      791640        748398        -5.46%
BenchmarkE2Mul-8        108           106           -1.30%
BenchmarkE2Square-8     67.5          65.9          -2.49%