feat: add audible request, for what the api command is not

[mkb79]

Based on #311. The diff below is against fix/api-takes-a-path, not
master. Both belong in the same release: #311 narrows api to the Audible
API and to JSON, and this is where everything else goes, so that nothing is
taken away without the replacement being there. Merge #311 first; GitHub
retargets this one.

audible api speaks to the API and expects JSON back. The rest of what
audible-cli reaches had nowhere to go from the command line: the website a
companion file comes from, the delivery host that hands out the audio, Amazon's
CDE service for annotations, and every answer that is not JSON.

audible request https://www.audible.de/companion-file/B07J2M2VC7 -o chapters.json
audible request https://cde-ta-g7g.amazon.com/FionaCDEServiceEngine/sidecar \
    -q "type=AUDI" -q "key=B07J2M2VC7" -i

Which hosts, and why there is a list at all

The request carries the credentials of the profile, so the list of allowed
hosts is the list of who may be handed them: api, www, cds and
api.amazon of any of the 11 marketplaces, plus cde-ta-g7g.amazon.com. That
is 45 hosts, derived in constants.py from the library's marketplace list, so
it grows with it. A url naming anything else is refused before any credentials
are loaded, as are plain http, a port of its own, and a name and password
written into the url.

What passes through untouched

The url is sent as it was written, not parsed and rebuilt. A delivery url is
signed over those exact bytes, and reading its query into pairs turns %FF
into the replacement character, %20 into + and flag into flag=.
--query appends to the raw query instead of merging into a parsed one.

The body goes out as it was given — bytes, not JSON — and the answer is
streamed, so a file of any size can go through --output without being held
whole in memory. The method is whatever the host understands, PATCH and
OPTIONS included.

--include/-i writes the status line and the response headers in front of the
body, wherever the body goes; --dump-header/-D writes them to a file of their
own. A header that came twice is written twice. A redirect is shown, not
followed. An error status ends the command after the answer has been written,
because the body of a refusal is what explains it.

Shared with api

QueryPair and HeaderPair move to _params.py, so both commands split a
query and a header the same way and refuse the same headers. One name differs:
api writes the body itself and always as JSON, so it owns content-type as
well, while here the body is raw and its type is the caller's to declare.

Tests

59, including two that go through the real AsyncClient behind a mock
transport, where a stand-in cannot help: the url that actually leaves, and that
a redirect is not followed — the second needs two hops to say anything.