[Nvidia] Implement pseudo-mersenne prime (Crandall Prime) acceleration

[mratsim]

Follow-up to #557 and #559, #445 needs to be ported to the Nvidia backend for proper Edwards25519 support on GPU (with raw memcpy instead of representation conversion).

We basically need to convert:

• https://github.com/mratsim/constantine/blob/v0.2.0/constantine/math/arithmetic/limbs_crandall.nim
• or if add-with-carry is supported: https://github.com/mratsim/constantine/blob/v0.2.0/constantine/math/arithmetic/assembly/limbs_asm_crandall_x86.nim