Support consuming IL(C#)-defined generic T with AllowByRefLike anti-constraint

[T-Gro]

Implements #16800
Follows https://github.com/dotnet/runtime/blob/main/docs/design/features/byreflike-generics.md
Builds upon: https://github.com/fsharp/fslang-design/blob/main/FSharp-4.5/FS-1053-span.md

Deserves a (small) RFC update: fsharp/fslang-design#787

[github-actions]

❗ Release notes required

---

✅ Found changes and release notes in following paths:

Warning

No PR link found in some release notes, please consider adding it.

Change path	Release notes path	Description
src/Compiler	docs/release-notes/.FSharp.Compiler.Service/9.0.100.md	No current pull request URL (#17597) found, please consider adding it

[vzarytovskii]

Shall we add tests for builders? Classic and resumable? Since a bunch of continuations are getting codegened there as well as inlining and state machines.

[T-Gro]

Shall we add tests for builders? Classic and resumable? Since a bunch of continuations are getting codegened there as well as inlining and state machines.

We don't propagate that generic anti-constraint to F# code, it really is only allowed for an ILCall expression.
Any specific idea in mind?

You mean like a generic builder, where the typar instantion would be e.g. MyBuilder<IEnumrable<Span<byte>>> ?
Or using a them inside a CE ?

[T-Gro]

I am thinking of some sensible real-world-like usecase involving a builder.
If anyone has an idea, I can definitely put it together as a test case (in this case, one we would fully run and let verify via the runtime)

[vzarytovskii]

Or using a them inside a CE ?

Mostly this, generic builders are less of a concern, I think.

I am thinking of some sensible real-world-like usecase involving a builder. If anyone has an idea, I can definitely put it together as a test case (in this case, one we would fully run and let verify via the runtime)

For example, if we have a collection builder and we define a bind or yieldfrom for some source with the type of (source: seq<ReadOnlySpan<int>>). I wonder what would we codegen in this case.