std::offload regression - requires lto=fat even though fat-lto is already enabled

When compiling for gpu offload, we always enable fat-lto:

[profile.release]
lto = "fat"
panic = "abort"

[profile.dev]
lto = "fat"
panic = "abort"

Lately I got hit with the following error when compiling a multi-crate project:
error: using the offload feature requires -C lto=fat

Commenting out the following lines solved it

-                if tcx.sess.lto() != rustc_session::config::Lto::Fat {
-                    let _ = tcx.dcx().emit_almost_fatal(OffloadWithoutFatLTO);
-                }

https://github.com/rust-lang/rust/blob/61d7280f3c4c63fa24c56bdaa9a446151b5a30dc/compiler/rustc_codegen_llvm/src/intrinsic.rs#L239

I expected to see this happen: It used to work

Version it worked on

It most recently worked on: Main, a week or two ago?

Version with regression

current main

cc @bjorn3 since you seem to be leading the lto cleanup efforts. Does my check strike you as something that checks for the wrong thing, or could it have regressed due to your changes? It's nothing urgent, I just disabled it locally till I have more time for debugging next week.

Relatedly, thanks for your cleanups! I wanted to give autodiff with thin-lto another try soonish, now that most of your changes landed.

<backtrace>

Start at compiler/rustc_codegen_llvm/src/intrinsic.rs around the offload and OffloadWithoutFatLTO check. Reproduce the multi-crate GPU offload build with the shown profile settings, then inspect how the session reports LTO mode. Done means a configuration with fat LTO no longer produces the erroneous requirement error.