Prior to this commit the `tests/tests.rs` integration tests had an
unconditional 1 second sleep waiting for a server to become available.
This is slightly problematic:

* Waiting a full second if the server becomes available sooner is
  unnecessarily slow.
* Conversely, we can't be sure the server is available after the 1s
  sleep.

This commit replaces the sleeps with a loop of up to 10 iterations that
tries to connect to the server port. If any iteration succeeds we return
immediately to avoid unnecessary waiting. If an iteration fails, we wait
longer, increasing the duration each time we wait (up to a maximum of 10
iteration).

Prior to this commit the `server()` test would call `assert_eq!` to
assert on client output matching expected _before_ calling `srv.kill()`
on the `Child` spawned by `server_command()`.

The upstream rustdocs for `std.process.Child` mention:
> There is no implementation of Drop for child processes, so if you do
> not ensure the Child has exited then it will continue to run, even
> after the Child handle to the child process has gone out of scope.

The net result is that if the assertion in `server()` fails, we leave
a running server instance behind. This manifests as a port conflict when
re-running the test without first killing the orphan.

This commit moves the `srv.kill()` call to before the `assert_eq`. This
avoids leaking the server process if the assertion fails.

This commit adds a few quality of life improvements to the server
integration test:

1. The `--silent` curl arg is removed so that we can get diagnostic info
   from curl on stderr in the event of a failure. Our test only asserts
   on stdout content, which remains unaffected by this change.
2. If the `curl` command invocation status isn't success, we print the
   stderr content enabled by not using `--silent`. This will provide
   more diagnostic context to investigate the failure.
3. The `assert_eq` is changed to assert on a string value created from
   stdout instead of raw bytes. This is easier for humans to diff in the
   test output compared to a slice of byte values.

Combined these changes should make it easier to diagnose future test
failures without needing to iterate in CI.

In the event the `server()` integration test fails it's helpful to know
the version of `curl` that was used. This commit includes this
information when the `curl` client invocation returns a non-zero status
code.