[tarcieri]

Move nested decoding, length bounds, completion checks, and failure state into a shared Decoder<R> type which owns the underlying reader.

This removes considerable complexity from the reader types, including their individual Reader::read_prefixed implementations and the supporting macro, leaving them responsible only for supplying bytes and reporting the remaining length.

Nested decoding now uses a common read_nested implementation which temporarily narrows the decoder's bounds. Each nesting level uses the same Decoder<R> type, preserving no_alloc support without the recursively nested generic reader types that previously caused monomorphization bloat.

This is a breaking change: Decode implementations now accept &mut Decoder<impl Reader>, which is slightly uglier from an API perspective but enables a much better underlying implementation.

A blanket From<R> implementation keeps call sites concise by allowing Readers to be converted into a Decoder with into()..

We are also motivated by #615 to make another breaking ssh-encoding release before the next ssh-key release: making the remaining infallible Mpint conversions fallible would allow its size limits to be enforced consistently.

Because this is a breaking change, the versions of ssh-cipher, ssh-derive, and ssh-encoding have all been bumped to v0.4.0-pre.

[tarcieri]

Sidebar: I think this sort of pattern might be interesting to adopt in der in the next breaking release