Speed up replace and canonicalize by improving type stability

[SebastianM-C]

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

This PR attempts to fix #2608 by avoiding boxing and runtime dispatch. So far this works for the case of homogeneous parameters, but for heterogeneous ones, we end up iterating over a tuple of different types, which is unstable.

[SebastianM-C]

I fixed a few more cases and added tests with JET and @inferred for replace/replace!. For canonicalize there are some cases which can't be inferred due to storing things in vectors, but that could potentially be addressed in the future with a vector type that stores its size in the type domain. I was a bit surprised that replace avoids the issue, but I think that the main factor is that we're effectively unrolling the iterations at compile time.

I think that this should not increase compile times as the tuples are used for the number of types, so we don't expect a ridiculous number of types in the portions.

[AayushSabharwal]

LGTM!