The `DecoderReader` used an internal buffer for reading which came with
a number of disadvantages, such as:

* Needless copying from in-memory readers (slices)
* Double-buffering already-buffered readers
* The bytes are lost when dropping or calling `into_inner`
* `BUF_SIZE` is not configurable
* `std::io::BufReader` has access to some unstable optimizations which
  this crate cannot use
* Reinvents the wheel; there already is `std::io::BufReader`

This change removes it and requires `BufRead` instead. Decoding is
implemented on top of `fill_buf` for larger chunks with fallback to
small, stack-allocated buffer for tiny chunks that may appear at
boundaries.

To resolve borrowing problems `decode_to_buf` had to be removed which
also enabled decoding bytes directly into internal buffer when the
buffer to be filled is small rather than into a temporary buffer which
is then copied.

This improves performance of reading by around 7-22% on my machine.