Try to hoist the conversion of diverging calls to aborts to a layer above MIR lowering

In https://github.com/rust-lang/rust/pull/122580 I added some code that prevents the magic compiler_builtins crate from linking against core. The replacement of calls happens as late as possible, so we end up with some odd behavior, where the LLVM IR and LLVM bitcode contain declarations of panic functions from core that are never called. We have those declarations because of this loop:
https://github.com/rust-lang/rust/blob/d6eb0f5a09247ff8443e68b4d17e0350c74bafb1/compiler/rustc_codegen_llvm/src/base.rs#L87-L89

We have MonoItems for the panic functions, and no calls to them. The key point is that we have MonoItems for items that we know would be invalid to call.

Is it possible to not have those MonoItems in builds of compiler_builtins?

One possible implementation of this would be to hoist this "error or abort on upstream call" to a MIR transform that's only run when building compiler_builtins. Then maybe we could drop the logic from cg_ssa?

Start by reading the MonoItem loop in compiler/rustc_codegen_llvm/src/base.rs and the existing "error or abort on upstream call" logic in cg_ssa. Trace how MIR lowering handles compiler_builtins, then determine whether that behavior can move to a MIR transform used only for compiler_builtins. Done means invalid panic functions no longer produce unused MonoItems or declarations in its LLVM IR and bitcode.