generated CodeInfo inlining should happen in Base rather than in Cassette

[jrevels]

To properly leverage JuliaLang/julia#22440, one needs to reimplement a bit of work that would normally be done as part of inlining during type inference.

Cassette currently does this as part of CodeInfo lookup, but this should really be in Base. Otherwise, the code will easily become stale w.r.t. inference changes. For example, JuliaLang/julia#22826 has already broken the current implementation in Cassette.

This pass could be either be exposed from reflection.jl (maybe as an inline keyword arg to lowered_code), or automagically performed by the compiler after generator expansion in the case where expansion yields a CodeInfo object instead of an Expr.

[jrevels]

After a bunch of discussion, it's becoming clear that this situation is unlikely to be resolved until Julia has a separately versioned compiler API, or something similar. I'm going to close this issue here, since at this point, it's not really a Cassette problem and more of a Julia problem.