This change is needed to allow platform and client specific
trampolines to modify the runtime behaviour of an endpoint, or even
just allow it to be accessed from outside the router.  Once an
endpoint goes into the router, getting it out is somewhat tricky.  It
is much easier however to just make an Arc<T> copy first, then passing
it into the Router.

Ideally there should be a second interface that endpoints need to
implement that allows destruction from externally, making the
remove_ep and get_ep APIs on the router redundant.

This commit changes a lot of things, but in a way represents a working
state interaction between the Kotlin (android) code and the Rust
native code.  Some issues are still present, such as having to
manually std::mem::forget(...) state types, instead of being
automagically forgotten.  But this can be fixed later.

Also, as this change set grows it's becoming obvious that the current
separation of android specifics, and java FFI handles in libqaul isn't
working very well.

When refactoring we should probably either move everything into the
android-support crate, assuming that the APIs of libqaul and similar
probably won't be consumed by other Java applications, or we should
make sure that there is enough tooling extrenally (rust4j for
example), to make it possible to have libqaul/ffi/java bindings that
handle everything from type conversions to java model declarations.

This allows the minimum SDK version to be dropped back down

We've had issues with libqaul::user_delete not completing when running
in parallel.  So far it's not clear to me why this is happening, and
I have yet to consistently reproduce it.  As a quick fix, limit the
test runs to 1, which seems to improve the rate at which it works.