Set environment variable for doctests

[CobaltCause]

Describe the problem you are trying to solve

Currently, it's possible to detect whether a procedural macro is being used in a normal context (i.e. library or binary crate) or integration test depending on whether the CARGO_TARGET_TMPDIR environment variable is set. This is very useful for proc-macro-crate, which allows usage of a proc macro using types from a crate in the crate defining those types. However, there appears to be no way to distinguish between being used in a "normal context" and being used in doctests.

Describe the solution you'd like

In order to allow this, I think either CARGO_TARGET_TMPDIR should be set for doctests if that makes any sense, or a new environment variable should be introduced that indicates that the current build is for a doctest.

Notes

See also this issue: bkchr/proc-macro-crate#10

[weihanglo]

I feel like CARGO_TARGET_TMPDIR just happened to be there. It wasn't designed for detecting compile mode. Setting CARGO_TARGET_TMPDIR for doctests seems like an anti-pattern to me, as doctests should be independent like unit tests as possible as they could. To be fair, I don't feel like crate compilation should be aware of any compiling context, except build scripts.

However, proc-macros are a bit special. People use it like a build script, so… not sure at all.

At this time, I would suggest using workspace and have a test suite member, like how serde is doing: https://github.com/serde-rs/serde/tree/7b548db91ed7da81a5c0ddbd6f6f21238aacfebe/test_suite

[CobaltCause]

Oh sorry, I actually meant to close this once I found rust-lang/rust#54363 but forgot about it. It seems to me like the solution proposed in rust-lang/rust#54363 (comment) might fix this problem? In any case I think that issue is better in that it's about solving the underlying problem instead of applying random bandaids like suggested here.