It has to be done prior 3.0.0 release, so shame on me.
Ruby 2.2 will be EOL soon, so I believe 4.x release should
drop Ruby 2.2 support too. :D I might be too radical today though.

This makes it possible to access the original body object.

Expose `HTTP::URI#query_values` & `HTTP::URI#query_values=` which
delegate to the underlying `Addressable::URI` instead of using
Addressable's `form_encode` method.

800c581 broke `readpartial` method: it returns same chunk twice :(

Ref #454

Currently for each chunk read from the socket a new string is being
allocated. This is not necessary, because we're feeding that string
object into the HTTP parser and never using it again.

So we initialize a buffer string and pass it in when reading chunks.
This will make each chunk be read into the same buffer object, no new
string will be allocated in that operation.

In my benchmark this halves the allocated memory when downloading
large response bodies.

Webmock triggers `body.each` twice upon inital request recording:

1. to perform request
2. to record request data into webmock request registry

But even without WebMock, generally Body#each must be replayable IMO.

Squashed commit of the following:

commit 15be2e59d4eb1b0d8c9bb6f1a8d2956421635ba0
Author: Yuri Smirnov <tycooon@yandex.ru>
Date:   Mon Feb 12 15:31:47 2018 +0300

    rename #prepare_request to #build_request

commit 9fab69049aef13edcdc8e3714618322f6d9c9e71
Author: Yuri Smirnov <tycooon@yandex.ru>
Date:   Mon Feb 12 13:36:37 2018 +0300

    extract prepare_request method

See: #458

This prevents the issue of encoding errors when trying to concat binary
bodies with this header string.

This should resolve #490