Our decoding code reads bytes in very small chunks. Which is not
efficient when dealing with the OS where the cost of a context switch is
significant. People could already buffer the data but it's easy to
forget it by accident.
This change requires the new `io::BufRead` trait instead of `io::Read`
in all bounds.
Code such as `Transaction::consensus_decode(&mut File::open(foo))` will
break after this is applied, uncovering the inefficiency.
This was originally Kix's work, done before we had the `io` crate.
Changes to `bitcoin` were originally his, any new mistakes are my own.
Changes to `io` are mine.
Co-developed-by: Martin Habovstiak <martin.habovstiak@gmail.com>
263a8b3603