Previously extractors worked directly on `Request<B>` which meant you
had to do weird tricks like `mem::take(req.headers_mut())` to get owned
parts of the request.
This changes that instead to use a new `RequestParts` type that have
methods to "take" each part of the request. Without having to do weird
tricks.
Also removed the need to have `B: Default` for body extractors.
This changes error model to actually allow errors. I think if we're going to use this for things like tonic's route we need a more flexible error handling model. The same `handle_error` adaptors are still there but services aren't required to have `Infallible` as their error type. The error type is simply propagated all the way through.
Wrapping everything in `crate::body::Either` wasn't actually necessary
ans probably causes large body types. You can instead box the bodies in
the leaf services.
Previously, when routing between one or two requests the two body types
would be merged by boxing them. This isn't ideal since it introduces a
layer indirection for each route.
We can't require the services to be routed between as not all services
use the same body type.
This changes that so it instead uses an `Either` enum that implements
`http_body::Body` if each variant does. Will reduce the overall
allocations and hopefully the compiler can optimize things if both
variants are the same.
David Pedersen