LGTM

I continue to have mixed feelings about this...

If this was mostly to replace File::open(T: &Path) and similar with wrapper functions that return a regular File but with an automatically created better messages then I'd feel a lot safer/better.
However, using fs_err:

1. seems to spread throughout our public API (because it introduces its own File wrapper)
2. its use already caught me off-guard and
3. it seems that this small rewrite potentially introduced a small scope(?)-related bug in src/dist/cache.rs, because we tried to work around the fact that tempfile works with regular std::fs::Path;
4. using the wrapper type and then converting to std::fs::Path boundary forces us to call into_parts().0 (I glanced at the API and didn't notice a better conversion API, sorry if there is one) which seems cryptic and which we could solve already with "proper" error "handling" calls like with_context(...)

Maybe we could introduce our local util wrapper that automatically only wraps the filename in the error variant only for the File::{open, create} APIs? That should cover most of the uses already and seems like a lower impact change, instead.