Support for no_std with a homespun/third-party Error trait

[mzabaluev]

I'd like to find an alternative to #64 that would not break compatibility for 1.0 users.

How about adding a non-default feature to thiserror that would, instead of std::error::Error, derive an impl of its workalike trait defined in the thiserror library crate, or possibly a reexport of a dusted off and properly released core-error, that would only depend on alloc?

Going further, there could be a feature to select a bare-bones Error trait that would shed downcast and anything else dependent on alloc.

This way, derive(Error) would not be a lie since the macro will derive an Error impl. The users would be free to ignore the impl and only use Display and From if that's what they are after.

[mzabaluev]

Cc @IcyDefiance @roblabla

[dtolnay]

This unfortunately does not sound promising to me, since I think it has most of the same failure modes as the old no_std PR. I'll close the issue but I would love to see a working PR if anyone comes up with some other approach.