feat(follow_redirect): Preserve extensions during redirection

[0x676e67]

Motivation

Extensions are passed during request redirection. Extensions that currently rely on http 1.x are now cloneable.

https://github.com/hyperium/http/blob/613d3d465f222666327d61f5971133cad155f190/src/extensions.rs#L35

Solution

Solve some scenarios where redirection requires passing Extensions

[0x676e67]

Is this approach justifiable?

[0x676e67]

Four hours ago, tracing-core updated its version and upgraded msrv. CI seems to have failed because of this.

[tesaguri]

Although I was not involved in #348, I guess it was a mere oversight that the design decision to drop extensions hadn't been revisited in the migration to http v1.

That said, I think it's a breaking change so the commit message should clearly make that clear, as required in the CONTRIBUTING.md. But, well, the contributing guidelines also prohibit one from rebasing the branch after opening a PR. I wonder how are non-conforming commit messages supposed to be handled then?

[jplatte]

Well, FWIW I didn't know this repo had a CONTRIBUTING.md 😅
@seanmonstar do you care about the stuff in there? In particular force-pushes to PRs?

[seanmonstar]

Eh, I don't know where all that came from, was it imported from Tokio?

I don't personally care if people force push PRs. I know some people like to see the incremental changes after review passes, but I always have to review the entire change before I click merge, so I don't care.

[jlizen]

@0x676e67 Sorry for the slow review...

I think that this was probably an oversight to omit, but there are security implications to turning it on. Extensions can leak sensitive information that previously was dropped. So certainly a breaking change.

We should also keep parity with the headers APIs and have FilterCredentials support for opt out via policy. It's a trickier problem than headers since mostly userspace types that we can't have a default excludelist for. I guess we could do some clever things if we have any types gated behind peer feature flags that we want to filter out.

Otherwise it seems like its mostly a user-provided filter. Probably something doable with macros to make it not-horrid to pass in a static list of type names/ids.

A simpler and non-breaking option would just be a opt-in API. (Long term I think it should be opt-out, but it can be a follow up)