Uninline the division in the AMD64 backend

[koute]

Currently the division/remainder instructions in the AMD64 backend are implemented by generating quite a lot of code inline (due to having to handle all of the corner cases of emulating RISC-V division), and considering that division is a relatively rare and expensive operation anyway it probably doesn't make sense to bloat the code with it (I'd like to have some upper bound on how much native code one PolkaVM instruction can generate, and this instruction bloats it a lot).

Convert div_rem in crates/polkavm/src/compiler/amd64.rs into a generic routine (you'll probably want two, one for div and one for rem) that will be emitted only once by the recompiler (e.g. similar to the emit_*_trampoline functions, just with a ret at the end), and strip down the div_rem to essentially only 1) push the arguments on the stack, 2) call the generic routine, 3) once the generic route returns move the result into the destination register.