refactor: based on #515 generify groth16 MPC setup for all curves, flatten packages+ refactor

[gbotrel]

Based on #515 (cherry picked the commits from @HSG88 );

Did mostly stylistic edits; flattened the package structure, avoided couple of allocs, and generify for all curves.

To be merged after #561 ; MPC setup is now gnark/backend/groth16/bn254/mpcsetup .

A future PR (cc @HSG88 ) would probably be nice to ensure APIs don't panic but return errors + struct and public methods are documented (you can test output with godoc) .

On the perf note, the individual scalar mul on affine coordinates are a clear bottleneck, maybe we should use jacobian? (+ optimize the memory usage of the mulGLV path cc @yelhousni in gnark-crypto) .

• groth16 MPC setup
• fix Z in preparePhase2
• parallelize coefficients multiplication
• parallelize scalar multiplication
• refactor: expose all typed backends in gnark/backend (moved from internal/)
• refactor: flatten mpc structure, idomify APIs
• test: ensure phase2 serialization is tested
• test: added reference benchmark
• refactor: setup -> mpcsetup
• refactor: move utils in mpcsetup; limit api surface
• refactor: minor code cleaning
• build: generify mpcsetup for all curves

[ivokub]

I think it looks good. Didn't go through the math exactly, hoping that @ThomasPiellard had a look.

Considering that the usage is more complex, I would recommend adding documentation example and a bit of package documentation. If sounds good, then can also do it myself.

Maybe would be good to add type-switched methods into backend/groth16?