[tiptenbrink]

The ureq host wrote a fresh template and a fresh buffer for every request, and built a new ureq agent each time. It encoded each request twice: once into an empty buffer to learn its size, and once for real. It now has a client per provider that keeps what a request can reuse.

• azure::Blobs and s3::Objects write the template of each kind of request (get, put, delete, list) once, in new. The list options are given there, since the list template depends on them.
• Every request is written into one buffer that the client keeps. grown_to_fit writes into it as it is, and grows it and writes again only on Error::Capacity. Once the buffer fits the longest request, each request is encoded once.
• A request is kept as a Placement while the buffer may grow, and read back with placement.request(&buf). That drops the borrow of the caller's content, so the puts now pass their content to send themselves.
• Each client sends with one agent, which keeps its connections.
• s3::Objects derives the SigV4 signing key on the first request of each UTC day and keeps it. A unit test checks that a kept key signs for the rest of its day and that the next day's key is derived again.

The examples and the loopback tests make a client once and call its methods. azure_put, which calls the proto directly, is unchanged.

Not in this PR: encoders that return a Placement themselves, which would remove the .map(|request| request.placement()) at each call.