[clang] [Modules] Don't fail when an unused textual header is missing. (PR #138227)

[Ilya Biryukov via cfe-commits]

ilya-biryukov wrote:

>  And I'm not sure such a use case is worth supporting.
>  But I understand that is only my own interpretation which can be incorrect. And I want to believe you have a better use case that doesn't rely on accessing private headers

The code I shared compiles with no error before this change and has errors now. It is okay to discuss if our use-case is "worth supporting", but we heavily rely on this behavior and this has already broken us and this puts us in a corner.

Could we revert this and have a discussion on how to satisfy both our use cases on equal grounds?
Is there any reason why this commit must be present and head and can't wait until we figure out a mutual way out of this?

https://github.com/llvm/llvm-project/pull/138227