Add the 3x GPU prover

[mrmr1993]

This PR

• pulls in the 3x GPU prover code (with changes to link it to OCaml, see this branch)
  • all of the new C++ and CUDA files are copied directly from there, the only changes should be to #include paths
  • some of the CMakeLists.txts, dune files and .ml files were changed to make things build
• restructures the build system so that all of the C++/CUDA code lives beneath src/camlsnark_c/libsnark-caml
• adds a meta-package to signal that CUDA is installed
  • this lives in ./snarky_cuda, can be installed with opam pin ./snarky_cuda from the project root

There is a test that can be run with dune exec src/camlsnark_c/groth16_gpu_prover/test/test.exe.

NOTE: This PR doesn't integrate the prover with the snarky API yet, but it should be possible to test it using generate_witness (from #322) to get the input and using the preprocessing and proving functions in Groth16_gpu_prover.
Update: This is no longer true, see the comment below.

[mrmr1993]

Updated to compute the FFT domains properly and to add a GPU example (run with make examples-gpu)

[mrmr1993]

This comment was supposed to be posted before, but I forgot to press the comment button 😑

Updated. This now includes a conditional module in snarky that extends a normal Snark_intf.Basic/Snark_intf.S to introduce the functions relevant to the GPU prover.

The basic structure now is:

• the C++ and CUDA code all lives beneath src/camlsnark_c/libsnark-caml
  • GPU code is conditional on CUDA support via the CMakeLists.txt
  • cuda_prover has the C++ bindings for OCaml and the toplevel CUDA code
  • cuda-fixnum has the CUDA multiexp functions
  • prover_reference_functions has the C++ functions from the reference implementation, tweaked to support the CUDA code and OCaml bindings
• there is a new 'library' snarky_cuda in the folder of the same name
  • this library is empty, designed to be opam pin-ed to tell snarky to rebuild with CUDA support
  • this is hidden from dune with data_only_dirs in the toplevel dune file, so that dune doesn't think it exists before it's pinned
  • it should fail to install (on Ubuntu and Debian) if cuda isn't installed
• there is a new stubs builder in src/camlsnark_c/dune for the GPU code
  • this builds the CUDA code from the 3x GPU prover and the C++ stubs to link it to OCaml
  • this is conditional on the presence of the snarky_cuda package in opam
• there is a new library groth16_gpu_prover in src/camlsnark_c/groth16_gpu_prover
  • this binds the stubs to OCaml functions (using the libffi ctypes API), using a functor over a partial Snarky.Libsnark to ensure that the opaque types (Field.t, G1.t, G2.t, Proof.t, etc.) match
  • this is also conditional on the presence of snarky_cuda ((optional) in dune parlance)
• the main snarky dune file (src/dune) has a new 'alternative dependency' clause that selects between a no_gpu.ml and gpu.ml file, depending on whether groth16_gpu_prover is present or not
  • these have to be explicitly omitted from the snarky library's modules to avoid unconditionally building them both (and failing when groth16_gpu_prover wasn't build)
• gpu.ml implements some functors that extend Snark_intf.Basic/Snark_intf.S with preprocessing and gpu-proving functions.

[imeckler]

Should this be deleted?

[mrmr1993]

I was keeping this around in case Hamish wanted to keep testing using his old set-up. Very happy to remove this if we think he should just be interfacing with the OCaml code (or something else) instead.