Nvidia remastered

[mratsim]

This refactors the Nvidia backend in preparation for hardware accelerated elliptic curves and fields.
The Nvidia backend has been broken following AMDGPU and x86 experiments in:

• AMDGPU JIT compiler #453
• LLVM: field addition with saturated fields #456

This:

• refactors the hello_world_nvidia so that it uses only low-level cuda primitives (but it may use high-level LLVM ones)
• reimplement the add/sub/mul using the new framework which should significantly reduce boilerplate.