#955 allowed passing `verbose=-1` to silence the initial server listening log, it seems appropriate that this setting should also silence the server closing log

…972)

Fixes #970.

The issue here is that our `access_threaded` function, and thread-specific
arrays, all relied on `Threads.nthreads()`, which by default only returns
the number of threads in the `:default` threadpool. In the new world where
the `:interactive` threadpool exists, there can also be threads there.

There is a new function `Threads.maxthreadid()` which helpfully tells us
the max id of any thread, regardless of threadpool. So the proposed fix
here is that if that function is defined, use that, otherwise fallback
to `Threads.nthreads()`.

* Allow more custom retry logic

Allow user to pass `retry_delays` with a custom `ExponentialBackoff` object.
Allow user to pass `retry_check` that is a function of the form:
`check(s, ex, req, resp) -> Bool` which will be called if the default
retry logic doesn't allow retrying.

* Update docs

* Fix setting of response body for StatusError exception

As reported in JuliaCloud/AWS.jl#593,
the recently refactored retry logic was failing to properly write
the response body when a `response_stream` was used and a `StatusError`
was thrown.

* Fix error still being thrown

* Add test and forcenew option to avoid socket closing race condition

* New retry_check should take response body as arg

I realized we can't mess with the Response body field directly
when using the new `retry_check` functionality because we might
clobber a user-provided `response_stream`, which we set as the
Response body field. Instead, it should just be an additional
arg to the `retry_check` function.

* Fix test

* update docs; pass resp as nothing if none

* Avoid creating additional string when writing headers

* Define our own simpler Version type

* Update src/Messages.jl

Co-authored-by: Jacob Quinn <quinn.jacobd@gmail.com>

* Update src/Messages.jl

Co-authored-by: Jacob Quinn <quinn.jacobd@gmail.com>

* Test we don't allocate

* Tidy tests

* Remove unused code

* Update src/clientlayers/MessageRequest.jl

* Write headers to temporary buffer first

- ...before flushing to socket, to avoid task switching
  mid writing to the socket itself.

* Remove unneeded temp buffer in Stream startwrite

* Make loopback test less flaky by using Event instead of sleep

* Don't connect a socket for Loopback test

- unnecessary and was leading to occasional DNS Errors

Co-authored-by: Jacob Quinn <quinn.jacobd@gmail.com>