publish: Avoid redundant buffer copies

[Turbo87]

Previously, we were buffering the incoming hyper::Body into a bytes::Bytes instance, which we then wrapped in a Cursor, which we then used as a Read implementation to fill a Vec<u8>. This was somewhat wasteful because it meant that we were buffering the tarball payload twice: once in the Bytes and once more in the Vec<u8>.

This PR changes the implementation to take more advantage of the Arc-like behavior of bytes::Bytes. We implement a split_body() function, which turns one Bytes instance into two Bytes, one for the JSON metadata, and one for the tarball payload. This will just shuffle around a few pointers underneath, without performing any additional memory allocation.

In this PR we also slightly refactor the S3 uploader implementation to not require a redundant seek operation, which makes the above implementation a bit easier.