add a bundle subcommand that packs an OTA bundle from a manifest

[joeferner]

Replaces the two Python packers that had been copied between projects and then edited apart. The container comes from rpi-loader-ota, so the packer and the firmware that installs a bundle are now the same code by construction rather than by anyone remembering.

A bundle.toml rather than arguments: what goes into a bundle is a property of the project, not of the invocation. Arguments could just about carry a kernel and one directory, which is what the scripts did; they cannot carry a settings file, a firmware blob and a certificate with a destination each, and the moment they try, a card's layout lives in a Makefile in a form nothing can check.

Directory sources are packed recursively now that a bundle can carry a path, sorted so the same tree always produces the same bytes, and dotfiles are skipped at every level.

--upload posts the result to a running board and prints what it says the update cost its card. Plain HTTP, and ureq with default features off, so no TLS reaches the dependency tree of a tool whose endpoint is a board on a local network.

The tests build real directory trees and check the result by parsing it with rpi-loader-ota, the same code the firmware runs: a failure there means the packer and the installer disagree, which is the one thing the crate exists to prevent.

Note that make package for the CLI cannot pass until rpi-loader-ota 0.1.0 is on crates.io. cargo strips the path dependency when packaging and resolves the version, which is what proves the published crate would build.