certain tests to run a bit longer to prevent unnecessary failures on
some systems.

returned, regardless of whether the connection is open.

`==() is invariant to it already, and the hashes of two "identical" URI
objects should ideally be equal as well.

Case in scheme, host and so on also can desync the raw input uri from
the parsed components.

We used to simply put the raw string input URI into the sprintf_cache
for the object. Unfortunately, since __hash() depends on the output of
_sprintf('s') while `==() does not, this lead to identical URIs with
differing hash values, as well as non-canonically-formatted string
representations of the URI returned from (string)uri and _sprintf('s').

This is not perfect (as we currently have no way of properly chaining
errors), but this should provide more helpful error messages to
developers who somehow manage to throw in the executor after finalising
the promise some other way.

There is arguably little reason to use the executor 'API' if not for it
doing this.

init_promise() creates a promise and gets its future. The
Concurrent.results() - empty test throws away both. This
caused the gc to generate complaints about broken promises.

Fixes erroneous testsuite output.

Do not complain if the originator for an asynchronous request has
gone away.

Do not output a misleading message about failing to read the UDP
packet or connection refused, if it is the user callback that fails.

Fixes #10057.

Improves performance for large inputs.

Fixes #10056.

Do not call callbacks while holding the mutex lock.

Fixes multiple testsuite failures.

The API for finalise() raced with on_{success,failure}(), as
new callbacks could get registered after finalise() was called,
but before the state was updated.

Fixes #10055.

Thanks to Chris Angelico <rosuav@gmail.com> for the report.

Don't store raw backtraces, they may contain references to stuff
that the user doesn't expect to get extra references.