There is a related thread.
How do the multipliers emitted by LLVM compare with the assembly in gmp? The implementation space for arbitrary-precision arithmetic is really broad (just see how much code is in gmp or flint to pick two examples!). Practically, you aren't going to get good performance at all sizes without reimplementing a serious chunk of gmp, which rustc developers would have to do from scratch because gmp is under a less permissive license. And as part of the standard library it would have to be supported on the exact same architectures rust itself is supported on.