Reduce number of CompletableFuture allocations during async invocation

The current asynchronous invocations allocate a new CompletableFuture instance (plus a UniHandle, and a ComposedAsyncInvocationStage) for every interceptor that calls compose() or the other InvocationState methods that add a callback.

We can reduce this by having a list of handlers in each invocation stage, and only creating a new CompletableFuture when we need to change the command.