More robust testing

rustdoc/mdbook only has basic support for testing code snippets. There are lots of complex examples in the anthology that can't be supported, doing things like testing errors (which rustdoc only supports unstably), using other crates, and building large examples from previous examples. I haven't looked close enough at the needs of the docs yet, but a primary purpose of this whole effort is to get docs tested, so the testing infrastructure needs to be beefed up.