Remove unnecessary method specializations for Numbers

[dennisYatunin]

There appear to be inference issues arising from unnecessary specializations on Number types in several RecursiveApply methods. Specifically, the specializations for rdiv (but not for rmul) are the source of the first bug described in #1983, and removing them fixes that error. Removing these specializations may also fix other inference problems for operations on complex fields (e.g., matrix fields of non-scalar values).

Several specializations of rmuladd are not removed in this PR because rmap does not currently support 3 input arguments (only 1 or 2). This can be fixed in the future, especially if we end up rewriting rmap in terms of unrolled_map.

• Code follows the style guidelines OR N/A.
• Unit tests are included OR N/A.
• Code is exercised in an integration test OR N/A.
• Documentation has been added/updated OR N/A.

[charleskawczynski]

This is my favorite kind of PR, better yet favorite way to fix inference failures: delete code. Fantastic.

[charleskawczynski]

Actually, @dennisYatunin, would you mind adding a JET test for the case that is breaking on the main branch?

[dennisYatunin]

I just added a new test_non_scalar_5 unit test, which divides a matrix field of Covariant12Vectors by a scalar. On my machine, this test fails with the Number specializations and passes without them. If the CPU and GPU tests on CI pass, I think this PR is good to merge in. Unfortunately, it doesn't appear to fix any of the soft fails we currently have in our CI, but it definitely fixes the first part of #1983.

[charleskawczynski]

Excellent, thanks!

[dennisYatunin]

Wow, looks like the CPU test on CI allocates 16 bytes for the in-place broadcast expression, even though the version I ran locally doesn't allocate anything. The only difference I can see is that I'm running Julia 1.10.4, while CI is running Julia 1.10.5. @charleskawczynski, I think this might be indicating another compiler regression?

Although the CPU test allocates, both it and the GPU test are able to run without crashing. So, this PR is definitely improving inference, and I'll merge it in with the CPU test as a soft-fail for now. We can look into the allocations issue in a separate PR.