No idea, but I can tell you how fast it would go in C using the GMP routines, as I benchmarked a lot of tweaks and misc. formulas.

Close to 690k jumps /s, in-place point addition, no reallocs - this with using lowest level mpn_* routines (assembler optimized).
Around 638k jumps/s with the mpz_* routines.

Compare this to using the routines in libsecp256k1 field_impl.h and same formula steps:
affine + affine: 852k jumps/s (1 inversion, 2 multiplications, 1 squaring)

libsecp256k1 jacobian + affine addition -> jacobian result:
7.5M jumps/s (8M 3S) - removed safety checks since no point is the infinity and neither can be the result)
But... non-deterministic, I struggled for weeks to find a way to use a J point represented in multiple different ways to produce a stable hash, even a single one bit 50% probability hash as a base for deterministic jump). Seems we can only compare two J points for equality or non-equality, comparison result can vary its sign due to Z scaling.

It doesn't matter what Rust compiles down to, it can never ever generate machine code that runs faster than what the lowest level assembler routines can handle.

So we either need lots of threads (GPU) or some special hardware to speed things up.